'A Chinese Open Model Beat Everyone' — Kimi K3, I Rolled the Tape to Check

The last two posts told the same story. In blog #21, Anthropic put a Mythos-class tier above Opus and locked the strongest model behind the government and a few partners. In blog #31, OpenAI released GPT-5.6 Sol only to government-approved partners two weeks later. The conclusion: the frontier is splitting into what you can buy and what’s locked away.

On July 16, China’s Moonshot moved the opposite way. Kimi K3 — at 2.8 trillion parameters, the largest open-weight model to date, with full weights promised for July 27. While the U.S. builds a wall, China routed around it by letting you download the biggest one.

And the week’s headline compressed to one line. “A Chinese open model beat both Fable 5 and GPT-5.6 Sol.”

Is that true? This post checks. The short answer: half true, half marketing. Let’s roll the tape on only the comparable numbers.

First, what happened

• Scale: 2.8T total parameters. MoE (Stable LatentMoE), activating 16 of 896 experts per token. Active parameter count undisclosed (the “A50B” you’ll see around is a community estimate back-derived from the expert ratio, not a spec).
• Attention: Kimi Delta Attention (KDA), hybrid linear attention. Instead of a full KV cache, it uses recurrent state to control what to keep and overwrite, sustaining the 1M-token context.
• Mode: Native vision, 1M-token context by default, maximum reasoning always on and non-disableable (131K default output, up to 1.04M).
• Price: $3 in / $15 out per million tokens (cache-hit input drops to $0.30). That’s identical to Claude Sonnet 5’s list price (though Sonnet 5 is on a $2/$10 promo through late August). Three to four times its predecessor K2.6 ($0.95/$4) — the priciest a Chinese lab has shipped — but cheaper than Opus 4.8 ($5/$25).
• Weights: Full release promised by July 27 (a deadline, not a date). Hugging Face is the expected venue but not officially named, and the license is unpublished (K2 used Modified MIT).

That’s the spec. Now let’s verify “beat everyone” in three pieces.

Check 1: on overall intelligence, it didn’t — it’s 3rd–4th

First, a trap. The SWE-bench / Terminal-Bench / GPQA numbers each lab publishes shouldn’t be compared to each other. They’re self-reported on different agent harnesses, so Moonshot’s 88 and Anthropic’s 88 aren’t the same test. Moonshot’s own comparison table mixes leaderboard numbers, self-run numbers, and internal numbers — and the ProgramBench maintainer publicly objected to its partial-credit averaging. The headlines blur all this into “K3 won.”

The one apples-to-apples measure — one party (Artificial Analysis) running every model on the same methodology — is the overall Intelligence Index. Here’s how it lands.

K3 is 3rd or 4th, depending on how you count. AA’s own launch article headlines “#3” (best configuration per model family — unrounded, K3’s 57.1 edges Gemini’s 57), while AA’s model page, which ranks per configuration, shows #4. Either way, Fable 5 and Sol sit above it. The “beat everyone” headline collapses here. On raw overall capability, K3 got close to the frontier — it isn’t the summit.

On agentic work (GDPval v2 Elo) the gap widens — Fable 5 is at 1,760 vs K3’s 1,668 (Opus 4.8 is 1,600). For long autonomous agent runs, Fable 5 is still clearly ahead.

Check 2: but there are three places it genuinely did

Here the other half of the headline comes alive. K3 didn’t beat “everyone,” but on three specific axes it genuinely took first place — and these happen to be the axes you hit most often in real work.

(1) Among models you can download, it’s the runaway #1.

Being 3rd–4th overall flips to: the strongest model whose weights you can hold. The three seats above it (Fable 5, Sol, Gemini) are all closed APIs, and the top ones (Mythos 5, Sol) are locked behind government gates. For a country, company, or lab that can’t get Sol, K3 isn’t a runner-up — it’s effectively the ceiling of what’s in reach. The open-weight floor rose to just under the frontier, which is the real reason for the DeepSeek-style market reaction.

(2) #1 on blind preference for front-end code — beating the closed leaders.

In Arena.ai’s Frontend Code Arena (WebDev), K3 ranks #1 — still holding it as of July 19, at Elo 1,679 over Fable 5’s 1,631 and Sol’s 1,618. That’s not a benchmark measuring correctness — it’s a blind comparison where developers see two outputs side by side, not knowing which model made which, and pick the better one. There, people chose K3’s front-end output over Fable 5’s and GPT-5.6 Sol’s more often (76% pairwise win rate). Two caveats: the board’s Elo margin of error is ±17, so #1 isn’t locked in, and outside front-end K3 ranks only #9 on the overall Text arena. So this is a win strictly for one job — not “this model is smarter” but “for this job, I’ll use this.” Still: a model that’s 3rd–4th overall beat the closed leaders on human preference for a specific job.

(3) Cheapest per task.

The per-token rate went up from K2.6, but what matters is the total cost to finish a task — and K3 uses 21% fewer output tokens than the prior generation. The result:

For the same job, K3 is nearly half of Opus 4.8. The gap to Sol, though, is one AA itself calls “similar” — within noise. So it’s genuinely #1 on value, but the model it crushed is Opus, not Sol. The bigger lever in practice is elsewhere: the 90% caching discount that drops cache-hit input to $0.30.

So “beat everyone” is wrong, but “beat open-weights, front-end, and value” is right. The headline inflated those three into “everyone.”

Check 3: where it didn’t win — and where to be careful

For balance, the other side. K3 has weaknesses the headline hides.

• Hallucinations went up. On AA-Omniscience, accuracy rose 33%→46%, but the hallucination rate rose too, 39%→51%. It gets more right and makes more up. For factual work (research, fact compilation), that’s a red flag.
• It’s slow. AA’s benchmark shows 62 tokens/sec, below the ~73 median for comparable reasoning models — and in the first week, practitioners measured the live API as low as 26–28 tokens/sec. With max reasoning always on, responses are long. Not for latency-sensitive real-time use.
• Science, facts, vision go to Gemini. Even within the same tier, K3 earns its points on coding/agentic, while Gemini 3.1 Pro leads on scientific knowledge, factual reliability, and visual reasoning.
• The distillation question. The community’s biggest debate in the back half of the week wasn’t benchmarks — it was reports of K3 identifying itself as Claude in conversation, which resurfaced Anthropic’s earlier accusation that Moonshot and other Chinese labs mass-harvested Claude outputs. Nothing is proven — web-data contamination or roleplay leakage explains it too. But if it turns out true, the “open-weights achievement” story itself wobbles. Watch this one.
• The conditions on “open.” You can’t get the weights yet (as of July 19) — the promise is “by July 27” — and even then, reality bites: even a 1.58-bit quant exceeds 512GB of RAM, so a realistic minimum is four workstation-class GPUs or a 1TB-RAM server running single-digit tokens/sec. The new architecture (KDA, Stable LatentMoE) isn’t in llama.cpp or Ollama yet, and you must reproduce a “harness contract” (reasoning-history preservation, tool-parsing conventions) to behave like the app. Open is not the same as cheap or free.

Bonus: the most interesting part isn’t the leaderboard — it’s the kernel story

The thing Moonshot showcased hardest wasn’t a ranking. It was one case study. They gave K3 a real kernel from their own training infrastructure (FLA Triton, production scale: 96 layers, 8192-dim) with one task: “make it faster without changing the numerical results.” K3 ran for 15 hours without human intervention, designed a new two-phase algorithm plus kernel fusion, and cut forward+backward from 283.6ms to 114.4ms — roughly 2.5x. Fable 5, given the same task, reached similar final performance, but K3 reportedly improved faster per iteration.

Why this matters: that kernel isn’t a demo problem — it’s Moonshot’s actual training code. A model that speeds up its own training infrastructure makes the next model cheaper and faster to train, and that model improves the infrastructure again — the opening of the recursive self-improvement loop from #22, put front and center by a commercial lab. All of it is self-reported by Moonshot, with no independent verification. Still, this is the first frontier-tier launch whose headline exhibit is “how fast it improves itself” rather than “how smart it is” — which may outlast the leaderboard talk.

In practice: what should you actually use?

The takeaway across the three checks: this isn’t the era of finding “the one strongest model,” it’s the era of attaching whichever model won for that job (an extension of #27). Concretely:

• Front-end/UI coding, high-volume value work → K3. #1 blind preference plus cheapest-per-task lands exactly here. But validate over the API first (pre-weights), measure whether it really wins on your workload, then talk self-hosting.
• Hard multi-step reasoning, long autonomous agents → still Fable 5 (or Sol). The GDPval Elo gap (1,760 vs 1,668) is where you feel it.
• Factual-critical work → watch K3’s hallucination rate (51%). Always wrap fact compilation and citation in a verification step.
• Migration, three rules: (1) don’t drop the reasoning history (reasoning_content) across multi-turn; (2) don’t switch models mid-conversation (it breaks the state contract); (3) you can’t turn thinking off, so brake with output caps and task scope. Shape tool calls to what K3 expects, per MCP tool design.

So, rereading the headline

Rolling the tape on “a Chinese open model beat everyone”:

• What’s wrong: it’s 3rd–4th on overall intelligence. Fable 5 and Sol are above it; it trails on agentic work, factual reliability, and speed. “Beat everyone” is spin — and Moonshot itself says K3 is “not at the frontier.”
• What’s right: #1 among downloadable models, #1 on front-end blind preference, #1 on cost per task. It took exactly the three axes you hit most in real work.

The market replayed exactly this. The day after launch (July 17), Taiwan’s market dropped 6% and Nvidia wobbled — then recovered most of it within the day. A DeepSeek rerun, except this time K3 is a premium-priced model, so the read flipped to “frontier-class AI still needs expensive compute” and the dip closed fast. In Washington, reports surfaced of a White House executive order on open-source AI under consideration; from the other side came the rebuttal that “a lagging player open-sourcing a non-frontier model is just rational strategy.” Panic or rebuttal, what everyone reacted to wasn’t the ranking — it was the fact this post keeps pointing at: the ceiling of what you can hold in your hands just rose.

In #21 and #31 the frontier split vertically — the shelf below you can buy, the top shelf locked behind the government. K3 drew a horizontal axis: the closed-API side, and the open side whose weights you can hold. And the open side’s floor rising to just under the frontier is the week’s real event. Strip out the inflated “everyone,” and that one fact is still big news.

Closing #31, I wrote the question had become “who can use the strongest model, and what will you assemble from the tier you were given.” K3 widens the answer by one slot — for the first time, that assembly list includes “a just-below-frontier model you can download onto your own servers.” Looking up at the #1 on the leaderboard matters — but so, now, does splitting by data who actually won for the job in front of you.

References: Kimi K3 official blog (Moonshot AI), Kimi K3 – Intelligence, Performance & Price (Artificial Analysis), Kimi K3 achieves #3 in the AA Intelligence Index (Artificial Analysis), Arena.ai WebDev leaderboard, Kimi’s open model K3 nears GPT-5.6 Sol and Fable 5 (The Decoder), Kimi K3, and what we can still learn from the pelican benchmark (Simon Willison), China’s Moonshot AI releases Kimi K3, the largest open-source model ever (VentureBeat), Moonshot releases 2.8-trillion-parameter Kimi K3 (Tom’s Hardware), Kimi K3: The 2.8T Model Promising Open Weights—and a Hidden Harness Contract (rohitai), AINews: Kimi K3 (latent.space), Markets experience new DeepSeek shock (Fortune), Kimi K3 is no reason for China panic (Transformer). Benchmarks and pricing are as published on 2026-07-19 (updated after first publication with the week’s external coverage and a fact-check pass) and may change with the July 27 weights/license release.