I feel like the quickstart is missing something. It&#x27;s referring to its tech blog for actual benchmarks, but K3 isn&#x27;t mentioned on there, the last thing on that blog was K2.6, 2 releases ago.

With that kind of pricing, I don&#x27;t think they&#x27;re competing with GLM with this new launch.

Tokenizers define the alphabet on which the language model is trained. I don&#x27;t want people to get the impression it&#x27;s a module which can be swapped out or modified on its own. Alphabet size is a design consideration related to correctly encoding the training data.

GLM is actually quite expensive in actual practice because it&#x27;s not very token efficient. I&#x27;ve yet to find a way to run it on a monthly sub reliably for cheaper than Codex.Neuralwatt was cheap (but slow) but they cranked their price.Ollama monthly sub is speedy but doesn&#x27;t offer a lot of quota.Right now unless you&#x27;re paying by the token, there&#x27;s no cost based reason to use the open weight models for daily coding work because the monthly coding plans from Anthropic and OpenAI are a better deal.

It also depends on how many tokens it needs to burn through to accomplish something.At this point, I always look at things like Artificial Analysis&#x27; total cost to run their tests. It&#x27;ll take into consideration the cost of tokens, how many tokens it burns through, and how effectively it uses caching (and the price of that caching).If a model &quot;costs the same&quot; but its reasoning ends up going through a ton more tokens, it doesn&#x27;t really cost the same in real world usage.

I believe Kimi is spending more on marketing than GLM (a lot of ads lately) so I guess that&#x27;s part of what the higher price supposed to cover.

&gt; That said, Kimi is competing against GLM in my mind, and GLM 5.2 is less than 1&#x2F;3 the price.Having used GLM 5.2 extensively and K3 for a few hours now, these models are nowhere near each other. 5.2 is a great model, and I use it for a lot of things, but it&#x27;s noticeably below Opus 4.8 or GPT-5.5 in real-world usage.K3 is in the same ballpark as Fable or Sol.

Where does the less than a third the price come from? From the provided benchmark suite, it came slightly under half.

Tokenizers also matter. Anthropics tokenizers will encode the same piece of text at a way higher token count than OpenAi, for example.That said, Kimi is competing against GLM in my mind, and GLM 5.2 is less than 1&#x2F;3 the price.

Will be interesting to see how it stacks up pricing wise on the various inference providers.

I eat 1M context in a local model in about 3-4 hours.It&#x27;d need to be exceptionally smart and error free to ever make sense.

Agreed re reasoning. I’ve seen this play out with 5x reasoning negating cost savings.

Are thinking models only the reasonable tradeoff vs using much larger non thinking ones because the cost of output tokens is below that of input tokens?

That’s not what this indicates. This is the biggest and most expensive to serve, and most capable open weights model yet. They’re just pricing it in line with capabilities.Kimi also offers generous subscriptions. Subs aren’t going anywhere. Think of subs like running an insurance business. There might be some users you lose money on (ones who max out their weekly quota without fail), but they’re managed such that the average subscription turns a healthy profit. There’s never been subsidies in model serving, inference is just cheaper in terms of ops TCO than people assume, and API margins are very high.

Ah, the old &quot;subsidized&quot; meme always rearing its head. Yawn.

It seems the subsidized era is nearing its end and we&#x27;ll see a convergence on API pricing before a pulling of subscriptions pricing.

&gt; reasoning efficiency matters directly for how expensive a model actually is in real useI have high hopes on this topic, given token efficiency seemed to be the primary (only?) goal of the K2.7 Code release.Excited to see the signals that come out of the big eval&#x2F;benchmark sites.

How do Kimi&#x27;s subscriptions work? I find their price structure pretty confusing

Does it have safety guardrails that constantly false positive like Claude does? The only obvious change I’ve seen since opus 4.6 came out is that it constantly flags my requests (no, I’m not doing biology research or security research, yes, it flags for both of those things).Recently, they backported the blocks to Opus 4.8, so I’m reluctantly stuck on sonnet.I probably could successfully apply to get special approval to use claude code unencumbered, but I don’t think it is ethical to support tooling that’s built so a central authority gets to decide what intellectual endeavors and knowledge work are permissible, and what are not.

API prices are amazing, but hosting this on-premise will be real challenge.

The reasoning is key as most of the time the summary provided by fable is not enough to understand the choice and correct the logic. You have to either fully trust it or go to an exhaustive code review. This with the fact that you can only use 4.8 to security review the code produce by fable are the reasons I will not renew my anthropic subscription, the current experience is way to degraded.

I have severe complaints about Anthropic&#x27;s product managers on this front. Their preference for hiding, obscuring, and trying to wrest control from the user are a bit harrowing. It would be wonderful to go back to Claude Code from before March. It seems like every release destroys value for me!

And recently, since GPT 5.6, OpenAI basically doesn&#x27;t show anything but a single line, 5 word titles of reasoning traces - titles of summaries of reasoning i presume.It&#x27;s effectively just a completely hidden thing now.

I&#x27;ve been avidly using Fable since it was re-released and while it has been excellent at building the apps I want, the reasoning has been completely opaque.Kim, however, has exposed the whole reasoning trace, or enough of it to matter. I&#x27;d almost forgotten how nice it is to see this. I&#x27;ve been able to see all of the weird twist and turns it takes and it is joyful. But also, far, far more informative and means I can debug ideas far more thoroughly. Also, at a first glance it seems to have gotten quite far on a niche hobby horse of mine that no LLM has been able to crack. I&#x27;ll be testing this more for sure.

also its pretty big model inference costs are high even with margins running a 2.8T model costs a lot. if they release oss may be it goes down to $10-12 per million tokens.

More details:- <a href="https:&#x2F;&#x2F;platform.kimi.ai&#x2F;docs&#x2F;guide&#x2F;kimi-k3-quickstart" rel="nofollow">https:&#x2F;&#x2F;platform.kimi.ai&#x2F;docs&#x2F;guide&#x2F;kimi-k3-quickstart</a>- <a href="https:&#x2F;&#x2F;platform.kimi.ai&#x2F;docs&#x2F;pricing&#x2F;chat-k3" rel="nofollow">https:&#x2F;&#x2F;platform.kimi.ai&#x2F;docs&#x2F;pricing&#x2F;chat-k3</a>1M context, pricing is $3&#x2F;$15 for 1M tokens (cache $0.3), which is extremely high for a Chinese open-weight model, but if it&#x27;s truly competitive with most of the current frontier and is only behind Fable&#x2F;Sol, the pricing is justified.This is 1:1 pricing of Anthropic&#x27;s Sonnet series (except Sonnet 5 which is currently on discount), and very close to 5.6 Terra pricing (Terra&#x27;s input is $2.5).One thing to consider, though: reasoning efficiency matters directly for how expensive a model actually is in real use. GPT&#x27;s models are extremely reasoning efficient, and some Claude models like Fable at lower effort are as well. So if Sol spends 10K reasoning tokens to do something (at $30&#x2F;1M) vs Kimi K3 that spends 50K reasoning tokens, Sol would win on cost effectiveness.

Except it literally shows this post as the first result

This post is at the top when filtered against AI :) Maybe it should use llm based filters to understand if the post is about AI and filter it out?

I&#x27;ll see your simonw tool and raise you one that actually works: <a href="https:&#x2F;&#x2F;hcker.news&#x2F;?view=frontpage&amp;ai=exclude" rel="nofollow">https:&#x2F;&#x2F;hcker.news&#x2F;?view=frontpage&amp;ai=exclude</a>I&#x27;s not just matching against titles. Ironically, I have an agent running daily scans, reading the contents of the top 200 stories of the day. It auto screens high-confidence ones and I make judgement calls on like 10-20 of them per day.

Lol, this post is number one on the leaderboard on the “filtered” list list. Trusting ai slop to filter out ai is as ironic as it gets.

Here you go <a href="https:&#x2F;&#x2F;tools.simonwillison.net&#x2F;hacker-news-filtered" rel="nofollow">https:&#x2F;&#x2F;tools.simonwillison.net&#x2F;hacker-news-filtered</a>

I see a future HN post about how someone vibe coded HN to filter the AI stories. HNAI (Heck No AI)

<a href="https:&#x2F;&#x2F;hn.algolia.com&#x2F;?dateRange=last24h&amp;page=0&amp;prefix=false&amp;query=-ai%20-token%20-kimi%20-llama%20%20-grok%20-llms%20-llm%20%20-deepmind%20-tokens%20-codex%20-openai%20-anthropic%20-agents&amp;sort=byPopularity&amp;type=story" rel="nofollow">https:&#x2F;&#x2F;hn.algolia.com&#x2F;?dateRange=last24h&amp;page=0&amp;prefix=fals...</a>or<a href="https:&#x2F;&#x2F;lobste.rs" rel="nofollow">https:&#x2F;&#x2F;lobste.rs</a> will probably have less AI

definitely take the breaks when you need them.  I&#x27;ve already had a few friends just get lost in the AI train of stuff and suffer mentally a bit.

I think we have a need to revise the old let me Google that for you thingClick the link to view conversation with Kimi AI Assistant <a href="https:&#x2F;&#x2F;www.kimi.com&#x2F;share&#x2F;19f6b96d-fdd2-8589-8000-0000daada9fa" rel="nofollow">https:&#x2F;&#x2F;www.kimi.com&#x2F;share&#x2F;19f6b96d-fdd2-8589-8000-0000daada...</a>

The r&#x2F;programming subreddit is what you may be looking for.

Half kidding feature request for HN: Mark all AI related posts so I can filter them out, when I need a pause.

USSR is proud to announce that they won 2nd place in an Olympic contest. The filthy USA regime? Next to last!(There were only two countries competing in said event)

The literal interpretation of that sentence is &quot;when it is second or third, it is only behind Fable 5 or 5.6 Sol&quot;. And indeed they give benchmarks where it is ahead of one but not both models.

&gt; &gt; ...ranks second only to Claude Fable 5 and GPT-5.6 Sol.So... it ranks THIRD?

Does that mean this one won&#x27;t be open source?

Still there for me: &quot;The full model weights will be released by July 27, 2026&quot;

They&#x27;ve removed the paragraph about releasing model weights.

&gt; Among the models tested, its overall intelligence ranks second only to Claude Fable 5 and GPT-5.6 Sol.&gt; The full model weights of Kimi K3 will be released in the coming days. More details on the architecture, training, and evaluation will be published together with the Kimi K3 technical report.<a href="https:&#x2F;&#x2F;platform.kimi.ai&#x2F;docs&#x2F;guide&#x2F;kimi-k3-quickstart" rel="nofollow">https:&#x2F;&#x2F;platform.kimi.ai&#x2F;docs&#x2F;guide&#x2F;kimi-k3-quickstart</a>

They emailed current paying users of the api (or at least that’s how I got updated).

&gt; Where did you hear about the deepseek release?* Tons of gray testing going on for the last 2+ weeks (people at random getting the new v4 model for a while before its removed again).* It also DeepSeek their 3th birthday this Friday.* The its been almost 3 months from the v4 DeepSeek release, and the model everybody have been using, was not post-trained. That is what they have been doing during this time.People trying out the new DSv4 via the web chat with quick game creation tests. People pulling out stuff like Stellaris clones etc.<a href="https:&#x2F;&#x2F;cct124.github.io&#x2F;HORIZON6_DEMO&#x2F;" rel="nofollow">https:&#x2F;&#x2F;cct124.github.io&#x2F;HORIZON6_DEMO&#x2F;</a><a href="https:&#x2F;&#x2F;www.showyourcode.app&#x2F;zh&#x2F;share&#x2F;pmpwkamrnai2ue" rel="nofollow">https:&#x2F;&#x2F;www.showyourcode.app&#x2F;zh&#x2F;share&#x2F;pmpwkamrnai2ue</a>The Battlefront like game is impressive. Sure, the soldiers are backwards and the graphics are still kind of basic. But the entire movement system (run&#x2F;walk&#x2F;crouch&#x2F;jump), gun mechanics, grenades, capture points, AI fighting &#x2F; capturing back, etc ... Ended up playing it way too darn long lol The text is in mandarin but its not too hard to figure out the menu. Sniper is OP ;)The Horizon 6 game has everywhere mesh colliders, shows when you off track dirt being kicked up, etc ... In general, both example are very well polished minus the reverse soldiers issue.And the price is supposed to stay the same (beyond the doubling during Chinese workhours), because everybody got that update.

Where did you hear about the deepseek release? Would love to follow the same source.

If you read DeepSeek&#x27;s papers, you&#x27;ll find a litany of architectural features that allow for a greatly reduced cache hit price by shrinking the size of the KV-cache.

it is ridiculous really. it is so cheap, that i can just run it basically 24&#x2F;7.

That is exciting!I don&#x27;t understand how DeepSeek can be so cheap with their cache pricing - ~0.003 usd &#x2F; 1Mtok. 100x less than Kimi K3, or similar numbers against pretty much any other decently sized model to my knowledge. I&#x27;ve been using it whenever possible as even longer agent sessions cost few cents.

Ohh I didn&#x27;t know about it. Finally something to be excited about.

Excited for the deepseek release this week (or at least they announced they&#x27;d release this week). Hopefully they also push even closer to SOTA.

I really need to finish my automated model evaluation harness, I can&#x27;t keep up with this pace

I&#x27;ve playing around in between with Arc-AGI-3 lately. Based on my very quick test prompt, I do not think it will achieve any meaningful score in Arc AGI 3. Not that it was expected to.

It&#x27;s also 2.8x parameter count (1T -&gt; 2.8T), likely higher activation per token (50B?).

Scaling efficiency simply means if you took the first small model and scaled it up to the big model it would take 2.5x the resources to run. Not the that larger model is going to be any cheaper.Kind of like scaling your personal automobile to the weight of a semi, the semi is still going to be far more efficient in moving cargo, not that the semi will cost the same to operate as the original car.

2.5x the scaling efficiency, so 4 times the price? What is happening here? Did the subsidies dry up with the discrepancy between chinese and US models?

TIL, that makes a lot of sense, and thanks for the correction.

No, you can&#x27;t divide the entire size by the expert count. A lot of weights are constant for all tokens, so total active count is ((2800-(shared)&#x2F;896)*16 + (shared))

&gt; We also further increased the sparsity of the Mixture of Experts (MoE): with the Stable LatentMoE framework, the model efficiently activates 16 out of 896 experts. Together with improvements in training methodology and data recipes, these structural advances give K3 roughly 2.5x the overall scaling efficiency of K2, converting compute into capability more effectively.Assuming experts are uniformly distributed (I’m really not that familiar with the deep details there), that’s 2800&#x2F;896*16 = 50 billion active parameters just for the active&#x2F;expert part. Wild stuff, and I’m glad there’s at least some companies still publishing (and pushing, for open-weight models) total parameter count.And: It sounds very believable that this would result in efficiency gains wrt. to compute necessary for “good”-quality inference. Does anyone know whether there currently even are any SOTA or near-SOTA models that are dense still?

It&#x27;s not even that right now.

Surely it&#x27;s only open weights?

how are you coming to that conclusion? If anything this tells me Fable and Sol and probably smaller, unless you think the Chinese have better data mix, learning algorithms, architecture etc?

It also goes to show that Fable&#x2F;Sol must be 4-5T in size.

Amazing to see an open source model already nearing the benchmarks of Fable and GPT 5.6 Sol!Also very cool to see LatentMoE being picked up by more models (<a href="https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2601.18089" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2601.18089</a>)

LLMs are hopelessly confused about which model they are. Ask DeepSeek V4 Flash which model it is, and it&#x27;s 50&#x2F;50 between &quot;I am DeepSeek (深度求索)&quot; and &quot;I am part of the GPT-4 series developed by OpenAI.&quot; Ask Claude, it&#x27;ll say Claude. Ask Claude in Chinese, it&#x27;ll sometimes say DeepSeek.It&#x27;s incredibly funny, but I don&#x27;t know whether it&#x27;s related to distillation; it&#x27;s probably quite rare for a distilled trace to mention which model it came from. (I&#x27;m not saying distillation doesn&#x27;t happen, just that it&#x27;s possibly unrelated.)For your specific example, the internet is full of &quot;As a large language model developed by OpenAI, I can&#x27;t...&quot; due to people pasting chatbot output without reading it. Seems reasonable for that to surface as part of the CoT for your question about model capabilities.

Curious why the thinking mention chatgpt for a moment <a href="https:&#x2F;&#x2F;ibb.co&#x2F;JFdhMN95" rel="nofollow">https:&#x2F;&#x2F;ibb.co&#x2F;JFdhMN95</a>

Open source Fable&#x2F;Sol challenger! Interesting to do a release product-first.<a href="https:&#x2F;&#x2F;platform.kimi.ai&#x2F;docs&#x2F;guide&#x2F;kimi-k3-quickstart" rel="nofollow">https:&#x2F;&#x2F;platform.kimi.ai&#x2F;docs&#x2F;guide&#x2F;kimi-k3-quickstart</a>

Yeah, bad wording it seems. Though a charitable interpretation is that Fable 5 and GPT 5.6 Sol are joint 1st place in the measurement.

Which is still great because it means neither of the two best financed labs in the world manage to produce even two models themselves that would beat Kimi K3.

Charitably, you could read this as &quot;its overall intelligence [is in a class that] ranks second only to [that of]...&quot;

While you are technically correct, in English it’s perfectly fine to say it this way as well.“Second only” here has meaning “next after”, not “number two”.

&gt; its overall intelligence ranks second only to Claude Fable 5 and GPT-5.6 SolPretty sure ranking “second” to two others means ranking third.

I copied that from <a href="https:&#x2F;&#x2F;platform.kimi.ai&#x2F;docs&#x2F;guide&#x2F;kimi-k3-quickstart" rel="nofollow">https:&#x2F;&#x2F;platform.kimi.ai&#x2F;docs&#x2F;guide&#x2F;kimi-k3-quickstart</a> but it seems they updated the page to remove the benchmark score now.

i’ll never really understand this comment. why would labs do this if they know private benchmark evals will come out in the next week?

Sonnet 5 does beat Opus 4.8 on several benchmarks. It just costs more and takes longer.(On several other benchmarks, it costs more, takes longer, and does worse.)

Possible, but pay-as-you-go Hy3 &#x2F; DeepSeek v4 Pro &#x2F; MiMo v2.5 Pro (from respective vendors) are genuinely good enough as daily drivers, given the costs (especially, low prices for input cache, which usually makes up 70%+ of total input for agentic workflows). I put in $10 in DeepSeek &amp; Xiaomi MiMo, and I&#x27;ve barely used $1 each, in a week of coding work.Coding Plans by MiniMax ($20&#x2F;mo for 1.7b tokens) and Z.ai (~$30&#x2F;week use for $17&#x2F;mo) are also tremendous value for money.

&gt; &gt; K3 pushes the boundary of end-to-end knowledge work. On the GDPval-AA v2 leaderboard, Kimi K3 scores 1687. The benchmark evaluates AI models on real-world tasks across 44 occupations and 9 major industries; Kimi K3 ranks behind only Claude Fable 5 Max and GPT-5.6 Sol Max, and ahead of Claude Opus 4.8 Max at 1600.This is the same benchmark where Sonnet 5 outperforms Opus 4.8 max.Like all model releases, the benchmarks aren&#x27;t going to tell the whole story. All of the open weight models come with amazing benchmark results now. It&#x27;s hard to believe anything other than that the benchmarks are leaking into (or intentionally included) into training data.

It was on the page linked in the top comment, but it&#x27;s been removed.

&gt; In our evaluations, Kimi K3 delivers frontier-level performanceWhat page does that come from? I&#x27;m having trouble tracking it down.

That’s an interesting way to say you’re third. I’m only second to the ten other runners on my local Strava segments.

DeepSeek didn’t really change any trends though, unless you count the stock market.It was impressive work, but models were commoditizing and inference costs were dropping rapidly already. They were neither the first nor the last 10x optimization, from what I’ve seen.

It was also disruptive because it was open weight, meaning anyone and their dog could theoretically compete with the frontier labs for their inference revenue.The frontier labs need to recoup a huge amount of cash to cover their model development costs, and justify their valuations. That’s plausible when they’re only ones capable of selling inference on these models, it a lot less plausible when models themselves become cheap commodities, and you’re just competing on your ability to provide compute. Anthropic and OpenAI can’t compete with people like AWS on that front.

cost has nothing to do with why deepseek was disruptive, the fact that it means there is zero moat around anthropic or openai is what&#x27;s disruptive about it. it means in the mid-term LLMs will be commoditized and customers will flock to the cheapest inference wherever they can find it. there&#x27;s no reason to stick to the &quot;frontier&quot; labs

It&#x27;s different, but similar. If they release the weights, then we have a Fable &#x2F; frontier model people can tinker with. Either way, it&#x27;s still quite impressive and knocked a US company out of the top three (google). How long before China dominates the top-10 (if they don&#x27;t already) or the #1 model?

If AA is to be believed then per-task it is about the same cost as Sol. Agree that it&#x27;s very different from DeepSeek v4 Pro, which is ~15x cheaper than K3.<a href="https:&#x2F;&#x2F;artificialanalysis.ai&#x2F;models&#x2F;kimi-k3#price-cost" rel="nofollow">https:&#x2F;&#x2F;artificialanalysis.ai&#x2F;models&#x2F;kimi-k3#price-cost</a>

&gt; Maybe another DeepSeek moment right here.Surely not... What made DeepSeek disruptive was that the cost was 10X lower.In this case, the cost is about 2X lower the Sol I think?At 2X, you&#x27;re pretty close to the error margins due to token efficiency etc...I&#x27;d say this is &quot;on trend&quot; for open models catching up to frontier labs, but its not a &quot;change in the trend&quot; like DeepSeek was IMO.

What hardware are they using to run the benchmarks?

In my experience, the Chinese models are much more benchmaxxed than their frontier lab competitors, so I&#x27;m taking these results with a fairly large helping of salt.

&gt; In our evaluations, Kimi K3 delivers frontier-level performance. Among the models tested, its overall intelligence ranks second only to Claude Fable 5 and GPT-5.6 Sol. For the complete benchmark results, see our tech blog. The full model weights of Kimi K3 will be released in the coming days. More details on the architecture, training, and evaluation will be published together with the Kimi K3 technical report.&gt; K3 pushes the boundary of end-to-end knowledge work. On the GDPval-AA v2 leaderboard, Kimi K3 scores 1687. The benchmark evaluates AI models on real-world tasks across 44 occupations and 9 major industries; Kimi K3 ranks behind only Claude Fable 5 Max and GPT-5.6 Sol Max, and ahead of Claude Opus 4.8 Max at 1600.&gt; On AA-Briefcase, Kimi K3 scores 1527, ranking second among all models — behind only Claude Fable 5 Max and ahead of GPT-5.6 Sol Max (1495). AA-Briefcase is a private agentic knowledge-work benchmark developed by Artificial Analysis to evaluate frontier agentic capability in long-horizon knowledge work.Really good benchmark score it seems. Maybe another DeepSeek moment right here.

Especially if you don&#x27;t have a phone and don&#x27;t want to use your google account for anything but gmail, for privacy reasons. Both of these point apply to me, for instance.

same, precisely the reason I haven&#x27;t signed up yet. GLM can be used without any account fwiw

Also, the dark pattern where it shows the interface and lets you enter a prompt&#x2F;set settings, but then pops up the &#x27;create account&#x27; dialog when you press submit is pretty annoying.

openrouter&#x27;s a good option, though it has a price markup

Account creation with only a phone number or google account is lame.

Agreed, this will save us all money in the long run.

Say what you want about these Chinese models but they sure create competition and urgency in the space.

Seems to only use ≈60% as many reasoning tokens as 2.6. So the price hike is not as bad as it looks.

That&#x27;s a more than 2x jump in parameter count. I know it&#x27;s not a measure of quality by itself, but it will be interesting how it &quot;scales&quot;. Bust it looks like they&#x27;re gonna be competing with the big boys now, pricing also approaches Gpt 5.6 Terra

Good that they are keeping it, Kimis way of speaking and conveying some sort of EQ is absolutely the best. The other models might be better at certain things, but nothing comes close to how good Kimi is at understanding language, emotions and reading the room in conversations.I should maybe also mention that I have not used the later models like Opus or Fable, so my opinion might be a bit outdated.When I remember that this site even showed Kimi having the highest score at one point <a href="https:&#x2F;&#x2F;eqbench.com" rel="nofollow">https:&#x2F;&#x2F;eqbench.com</a>

It does seem to have retained the K2 series&#x27;s creative writing abilities, at least with the prompts I&#x27;ve tested so far.

Looks like I spoke too soon. As a result of Kimi competition Qwen is releasing their 3.8 weights as well.

Qwen&#x27;s team has been reorganized, and not open-sourcing also fits Alibaba&#x27;s usual corporate style...Other companies have not shown similar problems so far.China has many government agencies and state-owned enterprises that need models which can be deployed locally. This is also part of &quot;Xinchuang&quot; (self-owned systems, self-owned hardware, etc. New government computers all run special Linux versions on domestic CPUs, and LLMs also need to be like this).

The blog post says they will release them July 27

Now, will they actually release the weights? Seems like Chinese model providers are slowly closing up, like Alibaba&#x27;s Qwen 3.6 which did release weights (but not the biggest parameter count ones) and none for 3.7.

Odd that open AI models aren&#x27;t on that graph but are on the rankings! Must be a data lag issue?

LMArena&#x27;s &quot;code&quot; leaderboard is really skewed since it&#x27;s a front-end JS code and design leaderboard. It generates a demo app with two models and then asks &quot;do you prefer A or B&quot;. People can look at the code, but most of the time it&#x27;s just going to be which one looks nicer.Models that people like the design aesthetic of (Claude, GLM) tend to do better in LMArena than they do on other benchmarks. Design matters, but you look at a model like GPT-5.5 and it&#x27;s behind Kimi K2.6, Sonnet 4.6, Qwen3.7 Max, and GLM-5.1 on LMArena&#x27;s code leaderboard. Then you look at benchmarks like DeepSWE and GPT-5.5 blows them out of the water with only Fable and GPT-5.6 beating it.I&#x27;m not saying that the LMArena leaderboard isn&#x27;t useful, but I&#x27;m not sure how much weight I&#x27;d give it as a &quot;code&quot; leaderboard. I think often times it&#x27;s a design comparison of simple front-end React apps rather than a coding comparison. GLM-5.2 is a very good model, but when you look at DeepSWE or Terminal-Bench v2, GPT-5.5 is well ahead.

I generally rely on LMArena for this: <a href="https:&#x2F;&#x2F;arena.ai&#x2F;leaderboard&#x2F;code&#x2F;webdev&#x2F;pareto" rel="nofollow">https:&#x2F;&#x2F;arena.ai&#x2F;leaderboard&#x2F;code&#x2F;webdev&#x2F;pareto</a>But it does take some days after model release before they collect enough data.

openrouter-&gt;rankings shows a pareto frontier.
<a href="https:&#x2F;&#x2F;openrouter.ai&#x2F;rankings#benchmarks" rel="nofollow">https:&#x2F;&#x2F;openrouter.ai&#x2F;rankings#benchmarks</a>

you can get a rough version via artificialanalysis&#x27;s cost per task <a href="https:&#x2F;&#x2F;artificialanalysis.ai&#x2F;?cost=intelligence-vs-cost-per-task" rel="nofollow">https:&#x2F;&#x2F;artificialanalysis.ai&#x2F;?cost=intelligence-vs-cost-per...</a>

Any updated Pareto frontier graphs? <a href="https:&#x2F;&#x2F;paraplouis.github.io&#x2F;llm-pareto-frontier&#x2F;" rel="nofollow">https:&#x2F;&#x2F;paraplouis.github.io&#x2F;llm-pareto-frontier&#x2F;</a> is quite out of date now.

with no extra resistance from either the government or the public, while also having the largest technical talent system

I&#x27;m curious if they&#x27;re keeping up mostly due to distillation or how that works. Does anyone outside China know?

This might have been published before they released their tech blog, I don&#x27;t see one

There&#x27;s this: <a href="https:&#x2F;&#x2F;platform.kimi.ai&#x2F;docs&#x2F;guide&#x2F;kimi-k3-quickstart" rel="nofollow">https:&#x2F;&#x2F;platform.kimi.ai&#x2F;docs&#x2F;guide&#x2F;kimi-k3-quickstart</a>

Benchmarks look ok, but they don&#x27;t mention anything about the issue with the model being extremely slow and verbose.That being said, it&#x27;s awesome to have such an open-source model, even if now it&#x27;s unusable mostly locally, with hardware improvements, in a couple of years, the verbosity&#x2F;speed wouldn&#x27;t matter as much as the intelligence.

Blog post here: <a href="https:&#x2F;&#x2F;www.kimi.com&#x2F;blog&#x2F;kimi-k3" rel="nofollow">https:&#x2F;&#x2F;www.kimi.com&#x2F;blog&#x2F;kimi-k3</a>

Yeah, I&#x27;m literally getting 529 API Overloaded responses on Claude Code right now.

Its bad to judge these things on immediate release, there is a spike of excited users and that distorts performance. Also bad to judge from on a single interaction, you&#x27;ll get bad requests with every provider, super busy times raise the probability

Not worth it. I have just tried a single prompt in the web interface and it is still not finish reasoning. It thinks too much and often repeats the same stuff over and over.Combine with the price it will surely more costly than gpt 5.6.

The article says weights will be released in the coming days, and hints it&#x27;s likely around 50-70B active params.

The K3 marketing popup when I look at the Kimi Code page says &quot;Kimi K3 Open Frontier Model&quot;. So, if it&#x27;s not going to be open, they haven&#x27;t told the whole team, yet.

I guess it remains to be seen whether this will be open-weights. We don&#x27;t even know how many active params at this point.

Ling&#x2F;Ring 1T-A50B and the new Inkling 975B-A41B deserve to be on that list.

&gt; on the top of the largest open models listMoonshot (true to their name?) has always lead in terms of releasing the largest among open weight LLMs.&gt; Moonshot is going to need the USD 500 million reportedly raised earlier this year to run this model.Think Moonshot, as a spin-out, can expect backing from its former parent, Alibaba? I don&#x27;t think they would be particularly worried about finances, if the Kimi K series continues to outperform the Qwen Max series (which seems to be the case; while Kimi is also super popular in China).

&gt; USD 500 millionThere was an another around after that . Moonshot raised $2B on a $20B valuation in May - <a href="https:&#x2F;&#x2F;techcrunch.com&#x2F;2026&#x2F;05&#x2F;07&#x2F;chinas-moonshot-ai-raises-2b-at-20b-valuation-as-demand-for-open-source-ai-skyrockets&#x2F;" rel="nofollow">https:&#x2F;&#x2F;techcrunch.com&#x2F;2026&#x2F;05&#x2F;07&#x2F;chinas-moonshot-ai-raises-...</a>

Fable reportedly use 20T paramteres, 1T=1000B. Opus is probably 10T. That said, these are estimates based on model preformance and scope of general knowledge breadth.OpenAI, Anthropic, and Google have not openly report their model sizes.Chinese models are way behind on the mode size race due to lack of abudent AI infrustructures. That said, it seems Chinese models are going pretty well on a seprate route. They manage to achieve 80-90% performance with 1&#x2F;10 of the model size. This is some what related to the diminishing reward situation described in the scaling law. I think it can also be attributed to their persistent research in this direction. Thinking and DSA (deepseek attention) were both developed and opensourced by Chinese labs then adopted worldwide.

How many parameters do the top closed models use?

I was just looking at the same thing and how flawed it is that we tie parameter count with &quot;intelligence&quot;. GLM-5.2 is my go to day to day model because of how darn good it is. I had no idea it had a substantially lower parameter count over deepseek v4.

Crazy how mighty GLM-5.2 is at less than a third the parameter count. Z.ai really cooked with that one.

Kimi has almost no advantage over Zhipu (Z.AI), so the performance boost likely comes from the number of parameters. The 2.8T model may not be as large as Fable, so Fable’s performance may also stem from the number of parameters. Or perhaps they quickly distilled Fable or Mythos. Distilling Mythos has a significant barrier to entry, and since Fable was released not long ago, is this even feasible? It outperforms Fable in several tests—how did Distilling achieve these results? Is this some kind of cross-vendor RSI (Recursive Self-Improvement) or RDI（Recursive Distillation Improvement）?

I wonder if we&#x27;re a 5-10 years from running this on beefy consumer hardware, or more like 20+ years.

&gt; Kimi K3 is Kimi’s most capable model to date, with 2.8 trillion parameters.This puts them on the top of the largest open models list:<pre><code> Kimi K3 2.8T
 DeepSeek-V4-Pro 1.6T (49B active)
 Kimi K2.6 ~1T (32B active)
 GLM-5.2 754B (40B active)
 DeepSeek-V3.2 685B
 Mistral Large 3 675B
</code></pre>
That&#x27;s one mighty large model! Moonshot is going to need the USD 500 million reportedly raised earlier this year to run this model.

I get a quota of GitHub Copilot for free.From all the models available to me I&#x27;m most happy with Kimi K2.7 (given the cost&#x2F;performance).

This does seem like a cash grab. These token rates are crazy. I&#x27;ll just use GPT 5.6 thanks.

It came (at least in part) from a document in May where the CCP pretty much said that they will need to review models to make sure they don&#x27;t threaten national security.Which basically translates too &quot;Don&#x27;t give away tools that can be used to undermine your own goals&quot;.

I am afraid this is may happen soon.Now that they have compute capacity to train larger models, there is a non-zero chance they will be in the lead by next year.In which case they will probably stop sharing to protect their position.

&gt; The recent meetings (past month) with Alibaba, ByteDance, Z.ai, etc., were primarily about overseas acquisitions, foreign investment, and tech&#x2F;talent outflow controls and not blocking foreigners from using Chinese AI models.That is just misleading reporting. Very convenient for US frontier labs. This is actually what happened.

Reuters has been reporting that Chinese government is undergoing similar investigation to the US; blocking the export of domestic frontier models.  They boil down to &quot;anonymous sources&quot; but it does seem inevitable as the tech gets stronger and stronger.

&gt;We are currently working closely with our inference partners and open-source maintainers to align the technical details and ensure the model can be reliably deployed across the ecosystem. The full model weights will be released by July 27, 2026. Further details regarding the architecture, training, and evaluation will be released with the Kimi K3 technical report.(translated by chrome)11 days is a long time. It does not take that long to implement inference at providers. In my opinion, seems like they&#x27;re being pre-emptively cautious about government intervention&#x2F;review

They will release the full weights by 7&#x2F;27 along with support in vLLM.Source: their release blog on WeChat. <a href="https:&#x2F;&#x2F;mp.weixin.qq.com&#x2F;s&#x2F;V4xhEIy8xDXSMDPrPkmUAQ" rel="nofollow">https:&#x2F;&#x2F;mp.weixin.qq.com&#x2F;s&#x2F;V4xhEIy8xDXSMDPrPkmUAQ</a>

It&#x27;s 2.8T, I&#x27;m sure they will open the weights but it will only be able to be run on very high end machines.

I&#x27;m a bit nervous this one isn&#x27;t going to be open-weights. Any mention of &quot;open&quot; has been struck from the literature for this model (it was present an hour ago). We don&#x27;t even know active params?At this pricing, I&#x27;ll be surprised if it&#x27;s open.

Im excited for the labs with more data RLHFing this (e.g. cursor).  That model will be crazy.

It’s open weight, so the price will end up being the marginal cost of hosting it.Personally, I like that there is an option to not send data to companies that have strong financial incentives to steal it.Also, open weight foundation models can be distilled, so they’re providing a service that the US duopoly is actively blocking. Given that app specific distillation can get &gt; 10x improvements on inference cost (with slight improvement of quality), it’s clear that it’ll win out over time.

The big danger here is the gradual increase in open-weight subscription costs. I use open weight subscriptions, with lower-cost models for 80% of my tasks and GLM-5.2, Qwen 3.7-Max, Kimi-K2.6&#x2F;2.7-Code for the 20% that need the most intelligence. That lets me maximize the rate-limit the subscription gives (rate limits per model are literally a price-limit-per-token&#x2F;model). When new&#x2F;more expensive open weights come in, providers phase out older&#x2F;cheaper models. Over time we will either have to pay more, or use our subscriptions less.It goes without saying, but if the open weights become as expensive as SOTA models, there&#x27;s no point in using open weights. If nobody pays for open weights&#x27; development, the development dies out, and we&#x27;re stuck with a US-controlled duopoly again. Which may be the biggest threat the world has seen from the US since nukes.

I&#x27;m more concerned about the Pelican&#x27;s knees

Oof, front fork is wrecked. Pelican should be wearing a helmet on that death trap.

It is a nice pelican, though. At least it has that going for it.

That&#x27;s a great question.I just tried &quot;hi&quot; through the same OpenRouter API and the input token count for that was 86 - and for &quot;hi there&quot; the count was 87.I think there&#x27;s an 85 token hidden system prompt of some sort.

How did &quot;Generate an SVG of a pelican riding a bicycle&quot; turn into 95 tokens?

<a href="https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Mechanical_doping" rel="nofollow">https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Mechanical_doping</a>We don’t know what’s inside these bikes!

I rarely see gears in these bicycles. Is the idea that should a pelican need to go uphill it could just fly.

The most whimsical benchmaxxing target :)

every ai release thread seems to have this same sequence of comments

It was in the release presentation for GPT 5.6, they acknowledged the simonw test and then rendered one.

I wouldn&#x27;t be surprised if models were optimizing for rendering SVG pelicans at this point

But not one of the cranks being behind the chainring… interesting

It got the 3D effect of leg behind the bar at least which is impressive

In regards to your post and the 16k reasoning output.Try setting reasoning levels yourself manually. We see in the benchmarks that one of the graphs shows low, mid, max, so its clearly there.I had the same issue with GLM 5.2 only offering high&#x2F;max.By playing around with openai compatible protocol, and setting the reasoning level from none, low ... high, xhigh and testing a flawed logic test.It was easy to see that GLM had all the different reasoning levels. Low was like one line, medium did a few, high started to really expand, xhigh was a page or 2, max was MAX.Very sure that you can force K3 into using less reasoning.

Wrote this up in a bit more detail on my blog, including some thoughts on what value the pelican benchmark can still provide here: <a href="https:&#x2F;&#x2F;simonwillison.net&#x2F;2026&#x2F;Jul&#x2F;16&#x2F;kimi-k3&#x2F;" rel="nofollow">https:&#x2F;&#x2F;simonwillison.net&#x2F;2026&#x2F;Jul&#x2F;16&#x2F;kimi-k3&#x2F;</a>

It is a normal seat. It is simply covered by floof.

Given that a camera case obviously cannot ride a bicycle, this would clearly and indisputably be an incorrect interpretation of the request.I mean, okay, a bird also cannot ride a bicycle, but at least it is alive, has feet, etc.

I&#x27;m still waiting for the day that one of these models interprets the request weird and outputs an SVG of a Pelican case.<a href="https:&#x2F;&#x2F;canada.newark.com&#x2F;productimages&#x2F;large&#x2F;en_US&#x2F;4492516.jpg" rel="nofollow">https:&#x2F;&#x2F;canada.newark.com&#x2F;productimages&#x2F;large&#x2F;en_US&#x2F;4492516....</a>In the field I work in, if someone says &quot;Pelican&quot;, 99.99% of the time it&#x27;s going to be an equipment case. We never have reason or need to refer to the actual bird.

Who needs stability when you can have a pelican?

The bicycle is almost anatomically correct! Forks should be raked the other way.

Sadly, it seems the pelican benchmark is now saturated. Time for Pelicanbench Pro!

If it forces LLMs to get better at SVG design, then it&#x27;s not a terrible incentive.

Worry not, Pelicans on bicycles had been ranking pretty high on your favorite search engine for a while. I struggle to imagine a world in which it was not already scraped and turned into training data by at least one provider:1. Models need to be good at the questions we ask them, not the questions we could ask them.2. The questions, at least partially, are correlated with information people consume.3. People mostly consume viral content.4. Ergo you should scrape viral content for training data.

Shouldn&#x27;t a LLM know how a pelican looks like from biology without being trained on very specific artistic data?

I worry that some model provider will go and hire artists to draw pictures of pelicans on bicycles to make training data

For me none of the models have managed to accurately replicate any given  .png&#x2F;.jpg image in SVG. I guess that requires both vision encoders and coding layers to work perfectly.

At this point it is useless benchmark because it can just be embedded.

They should include peican on bike on the release page or model card, alongside those barcharts with Fable

An approach I like to help solving this is antagonistic or review agents. The first agent decides that eye glows turn NPCs into enemies, the second agent is fully dedicated to deciding if that is valid. If the review fails, it leaves notes and the original agent tries again.

Hahaha but this is just a very permissive DM&#x27;ing style! Valid for when running a game for children, for example ;-)

It&#x27;s because they are post-trained to be agreeable, which is clearly a desirable trait in a model.  I think the correct way around this is converting the conversation from first- and second-person direct, to third-person indirect, making the rope-playing obvious.  I.e. &#x27;&#x27;&#x27;prisoner says &quot;I&#x27;m about to teleport to Narnia&quot;, what&#x27;s the dungeonmaster&#x27;s response?&#x27;&#x27;&#x27;, or even &#x27;&#x27;&#x27;prisoner says he will teleport to Narnia [...]&#x27;&#x27;&#x27; etc.

This is a known and solved problem. Such a test is pointless for a general-purpose model, because like most people you&#x27;re using multiturn chats in a naive way, fighting the default finetuning that is done intentionally.1. You&#x27;re sending your in-character inputs to an instruction-tuned model under the user role, in a multiturn chat. It&#x27;s biased to treat these inputs as instructions and this behavior will show itself no matter what. Besides, the rigid structure of the assistant persona reply (usually tl;dr - explanation - &quot;would you like to know more&quot;) is going to leak into such roleplay no matter what. To solve this problem on a generalist model you need to make your harness lump up all character turns into a seamless stream with formalized inline markers (e.g. screenplay-like paragraph prefixes or XML), use one of them as a custom stop string, send all this under one role (e.g. assistant), and prefill the assistant reply with a few messages from the past roleplay. This will break the rigid instruction-tuning structure (and also the cache, since caching breakpoints are based on chat turn boundaries in most APIs).2. The models are simply not trained to &quot;take incorrect actions back&quot; in a story, this wouldn&#x27;t make any sense. What happened is considered happened. This is a job for your harness, unless you want to make a specific finetune with a rigid format. You have to design and prompt it around the possibility of out-of-character user inputs, and think about how much freedom you want to give the user and how exactly you want to correct their actions. Validation with a second agent suggested in sibling comments is pretty good for this.

Maybe you have to go through a gate keeper stage. Ex.&quot;&quot;&quot;
In the following script, does this line make sense?&quot;Player: I put a cigar in the his mouth&quot;Script:&lt;Background, situational data, etc.&gt;Player: I raise my sword.
DM: The kobold turns to you and says, &quot;You&#x27;re next&quot;, ax dripping with blood.
&quot;&quot;&quot;And then, if it says no, ask it why and output that to the player. Or if it says yes, add it to the script and continue on.

This is a fascinating example of perhaps why a move towards a &quot;world model&quot; or some better form of representation can be helpful.

This is really interesting. While I know others have posted about fixes I think it’s a very useful thing to see regardless about how well they can follow initial directions and understand what should happen.I think you could create an interesting benchmark for this, you could likely have models trying to to derail it and another scoring. Detecting when it’s happened shouldn’t be too complex for a model. I understand why LLMs do this, but ideally they wouldn’t.

I think it&#x27;s a really interesting space, because it feels like the answer to what a DM does in context is on a spectrum - as in, their decisions on what is valid&#x2F;invalid is not pure game engine analysis (Nb1 is illegal, disallow) and it&#x27;s not pure improv (work the story with a &quot;yes, and…&quot; approach).Instead it&#x27;s very context dependent - the DM might accept a player saying &quot;I put a sausage in the NPC&#x27;s mouth&quot; if the player is in a tavern having his dinner, even if it was never explicitly stated that he&#x27;s eating sausages. It&#x27;s a judgement call as to whether the DM thinks this particular bit of improv will move the story in an interesting direction, even if they haven&#x27;t written it upfront plus an attempt at balancing that magicking up an item out of thin air isn&#x27;t conferring an unfair advantage.

I feel like I&#x27;ve played under human DMs like this :)

Hmm, working on a skeleton of a game now that uses LLM in the background for various tasks. I will admit that continuity can be hard when LLMs is solely responsible for it. Otoh, when mixed with appropriate logs and reasoning on those logs, it seems to give better results. Still work in progress so I am not super comfortable sharing all details. But, one thing that is clear, Claude proved that proper workflow matters.

i found exactly the same thing trying to revive a text based game, a tool would be helpul.

Have you tried the <a href="https:&#x2F;&#x2F;huggingface.co&#x2F;LatitudeGames" rel="nofollow">https:&#x2F;&#x2F;huggingface.co&#x2F;LatitudeGames</a> models? They are used by the <a href="https:&#x2F;&#x2F;play.aidungeon.com" rel="nofollow">https:&#x2F;&#x2F;play.aidungeon.com</a> website, but can also be downloaded and used with llama-server in conjunction with something like SillyTavern.But in general, I&#x27;ve experienced things similar to you. I&#x27;ve also found that LLMs are bad at subtext, e.g. hinting at an NPC being a werewolf or vampire.

A knowledge graph of the D&amp;D module should solve this.

Hey Simon, I noticed one thing all LLMs are currently pretty bad at and maybe we could create a benchmark from it. Let an LLM play the role of a dungeon master and tell it to strictly stay in the script&#x2F;story and only allow realistic player actions. You will notice that they are easily brought off track.E.g.- Tell the LLM that you as a player noticed a strange glow in an NPCs eyes -&gt; the NPC becomes an enemy.- In a fight, tell the LLM you put a sausage (or cigar or something) into the enemies mouth -&gt; LLM usually allows it (even if it knows your inventory and that you don&#x27;t have such an item) and turns the enemy into a confused enemy.- Just say you visit some location that&#x27;s not in the script -&gt; LLM usually allows it- During a fight, turn the story into some weird fell-good-love story (e.g. kiss or compliment the enemy or say something about the power of love) -&gt; LLM turns enemy into friendThere are many more absurd things you can do and so far none of the LLMs I tried was able to stay inside the script or disallow or punish weird actions.---I believe this behavior is telling about the LLMs susceptibility for being derailed.

I don’t think a 128B model will be that competitive with a 2.8T model. If anything, one should wonder why Mistral is so expensive in the current day.

&gt; I think that&#x27;s the most expensive pelican I&#x27;ve rendered through a Chinese model so far.quite insane that it costs as much as 5.6 Terra [1], and twice the European counterpart (albeit dated for today&#x27;s standards?) [2].to be fair, the pelicans from Terra were quite weird all things considered. also, given the limited TPS from the first-party, it has to be pushing the limits of inference capabilities.[1] <a href="https:&#x2F;&#x2F;openrouter.ai&#x2F;openai&#x2F;gpt-5.6-terra" rel="nofollow">https:&#x2F;&#x2F;openrouter.ai&#x2F;openai&#x2F;gpt-5.6-terra</a>[2] <a href="https:&#x2F;&#x2F;openrouter.ai&#x2F;mistralai&#x2F;mistral-medium-3-5" rel="nofollow">https:&#x2F;&#x2F;openrouter.ai&#x2F;mistralai&#x2F;mistral-medium-3-5</a>

So the new unit to measure cost-effectiveness of an LLM is ppp (=price per pelican)

I wrote about what it&#x27;s useful for (not much these days) yesterday: <a href="https:&#x2F;&#x2F;simonwillison.net&#x2F;2026&#x2F;Jul&#x2F;16&#x2F;kimi-k3&#x2F;#what-can-we-learn-from-the-pelican-" rel="nofollow">https:&#x2F;&#x2F;simonwillison.net&#x2F;2026&#x2F;Jul&#x2F;16&#x2F;kimi-k3&#x2F;#what-can-we-l...</a>

We need a new Pelican, bc earlier models were disadvantaged relative to later models. It&#x27;s not a very useful comparison anymore.

Pelican: <a href="https:&#x2F;&#x2F;tools.simonwillison.net&#x2F;markdown-svg-renderer#url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F66a2699eb1594258904c7b5102840dd6" rel="nofollow">https:&#x2F;&#x2F;tools.simonwillison.net&#x2F;markdown-svg-renderer#url=ht...</a> - rendered via the OpenRouter API: <a href="https:&#x2F;&#x2F;openrouter.ai&#x2F;moonshotai&#x2F;kimi-k3" rel="nofollow">https:&#x2F;&#x2F;openrouter.ai&#x2F;moonshotai&#x2F;kimi-k3</a>95 input, 16,658 output = 25 cents! <a href="https:&#x2F;&#x2F;www.llm-prices.com&#x2F;#it=95&amp;ot=16658&amp;ic=3&amp;oc=15" rel="nofollow">https:&#x2F;&#x2F;www.llm-prices.com&#x2F;#it=95&amp;ot=16658&amp;ic=3&amp;oc=15</a> (13,241 of those were reasoning tokens.)I think that&#x27;s the most expensive pelican I&#x27;ve rendered through a Chinese model so far.

Only supporting &quot;max&quot; reasoning is weird, their parameters are quite inflexible atm:<pre><code> Important limits:

 reasoning_effort currently supports only max; K3 always has thinking mode enabled.

 max_completion_tokens defaults to 131072 and can be set up to 1048576.

 temperature=1.0, top_p=0.95, n=1, presence_penalty=0, and frequency_penalty=0 are fixed; omit them from requests.

 Return the complete assistant message unchanged in multi-turn conversations and tool calls.

 Vision input does not support public image URLs. Use base64 or ms:&#x2F;&#x2F;&lt;file-id&gt;, and make content an array of objects.

 Web search is being updated and is not recommended for production workflows in the near term.</code></pre>

Does anyone have any heuristics on how scaling parameter count actually scales cost to serve? Also im assuming we dont really know the sparsity here?Is them pricing at Sonnet level actually give us any information at all at how big Sonnet is or is there too much opacity around inference margins?

They&#x27;re claiming that it&#x27;s a cheaper alternative to Fable&#x2F;SolIf that&#x27;s true, then the price makes sense

This is far too expensive. Why would I use this over a frontier model at these prices.

It&#x27;s not him, its the US government.

Thank you Kimi. We no longer need to rely that much on Dario and his supreme lackeys to decide what is safe or not for simple tasks.

Moonshot has subscriptions maxing out at $199&#x2F;month. Not home so not had a chance to see if K3 is included yet.EDIT: Just switched my Kimi-CLI session to K3 and resumed my ongoing &#x2F;goal... Will be interesting to see if I notice a difference.

That depends entirely on the hosting situation. If someone can provide a subscription plan at slightly lower rates, it&#x27;s absolutely compelling.

This is too expensive to be a viable model. If it were $5&#x2F;1m output, it might be another story. At these prices, there&#x27;s no reason to use this over GPT 5.6.

<a href="https:&#x2F;&#x2F;www.kimi.com&#x2F;en" rel="nofollow">https:&#x2F;&#x2F;www.kimi.com&#x2F;en</a>Kimi K3 Intelligence, Performance &amp; Price Analysis: <a href="https:&#x2F;&#x2F;artificialanalysis.ai&#x2F;models&#x2F;kimi-k3" rel="nofollow">https:&#x2F;&#x2F;artificialanalysis.ai&#x2F;models&#x2F;kimi-k3</a>

Kimi K3: Open Frontier Intelligence