|
ResearchAudio.io
Moonshot's 2.8T Model Admits Second Place
16 of 896 experts, 1M context, Sonnet pricing. The footnotes tell the real story.
July 18, 2026 · Frontier AI, decoded for engineers
|
|
2.8T
total parameters, the largest open model ever
|
16/896
experts active per token
|
1M
token context, 1,048,576 exactly
|
|
|
The largest open-weight model ever released opens its launch post with a concession. Moonshot shipped Kimi K3 on July 16 and stated up front that it still trails Claude Fable 5 and GPT-5.6 Sol overall. Labs do not usually write that sentence in their own announcement.
That candor made me read the rest of the post differently, footnotes included. And the footnotes are where this release gets interesting for anyone shipping agents on frontier models.
|
|
the lead
What 2.8 trillion parameters gets you
Moonshot calls K3 "the world's first open 3T-class model." Under the hood it is a sparse mixture of experts built on a framework it names Stable LatentMoE: 896 experts, with 16 firing per token. Input is text plus images, output is text, and reasoning is always on, with a single effort level (max) available for the moment.
Two architectural pieces carry the release. Kimi Delta Attention (KDA) is a hybrid linear attention design that keeps the 1M window tractable. Attention Residuals (AttnRes) replace standard residual connections, selectively retrieving representations across depth instead of accumulating them uniformly. Moonshot credits the pair, plus refined training recipes, with roughly a 2.5x gain in scaling efficiency over Kimi K2.
|
|
how 2.8T stays servable
|
|
1M context
text + images in, one long window
|
|
→ |
|
KDA
hybrid linear attn keeps long context tractable
|
|
→ |
|
16 of 896
experts fire per token (Stable LatentMoE)
|
|
→ |
|
AttnRes
retrieves across depth, not uniform residuals
|
|
|
|
MXFP4 weights + MXFP8 activations from the supervised fine-tuning stage on · claimed 2.5x scaling efficiency vs K2
|
Source: Moonshot AI, Kimi K3 launch post, July 16, 2026
|
|
|
The serving details matter as much as the math. K3 is quantization-aware trained, so the MXFP4 checkpoint is the model, not an afterthought. KDA breaks conventional prefix caching, and Moonshot has contributed a KDA cache implementation to vLLM that ships alongside the weights. Recommended deployment is a pod of 64 or more accelerators.
|
|
the scoreboard
Where K3 leads, where it trails
K3's launch table, run at effort max with temperature 1.0 and top-p 1.0, shows a clear shape. It leads on the long, grinding work: SWE Marathon at 42.0 against Fable 5's 35.0 and GPT-5.6 Sol's 39.0, plus top marks among the listed models on BrowseComp (91.2), DeepSearchQA (95.0 F1), Automation Bench, SpreadsheetBench 2, and OmniDocBench.
It trails where depth compounds. FrontierSWE goes to Fable 5, 86.6 to 81.2. Humanity's Last Exam without tools goes 53.3 to 43.5. On GDPval-AA v2, an Elo measure of economically valuable work, K3 posts 1668 against Fable 5's 1760 and Sol's 1748.
|
|
four rows that define the launch
K3 ■ Fable 5 ■ GPT-5.6 Sol ■
|
|
SWE Marathon (long engineering sessions · K3 ran in Claude Code)
|
|
Terminal Bench 2.1 (K3 in KimiCode · rivals at their own best harness)
|
|
FrontierSWE (dominance score · recomputed July 16)
|
|
Humanity's Last Exam (full set, no tools)
Source: Moonshot AI launch table, effort max, temperature 1.0, top-p 1.0
|
|
|
read the footnotes
The harness is part of the product
Moonshot documents its methodology openly, and the methodology changes the reading. K3 was scored inside KimiCode on most coding rows, while rivals got their strongest score across Terminus 2, Claude Code, or Codex. On Terminal Bench, the independent board's leading published score sits at 84.6, so K3's 88.3 in its own harness is notable but not yet an apples-to-apples result.
SWE Marathon is the exception, and that makes it the number I trust most. K3, Fable 5, and Opus 4.8 all ran inside Claude Code there, and K3 still led by seven points. Its widest coding win came inside Anthropic's own harness, not Moonshot's.
The footnote that stopped me: on PostTrain Bench, whenever Fable 5 refused a task on usage-policy grounds under Claude Code, the run fell back to Opus 4.8. Even refusal handling is documented. BrowseComp gets similar care: the 91.2 uses context compaction triggered at 300K tokens, the strategy from Claude's model cards, and the raw 1M-context run with no management still lands at 90.4.
Independent signals back the shape. Artificial Analysis scores K3 at 57 on its Intelligence Index, behind Fable 5 and the GPT-5.6 Sol settings and ahead of Opus 4.8, GPT-5.5, and GLM-5.2. And on the Frontend Code Arena, which is blind human votes rather than any harness, K3 sits first at 1679 Elo, ahead of Fable 5 and up from K2.6's 18th place.
|
|
Key insight: match the harness before you trust the number. If your agents run in Claude Code or Codex, the SWE Marathon row tells you more than any KimiCode-measured score, because it was measured where you actually work.
|
|
|
the economics
Here's the part nobody's talking about
K3 costs $3 per million input tokens and $15 per million output, with cached input at $0.30, a 90 percent reduction. That is Claude Sonnet pricing, and the priciest model any Chinese lab has shipped. K2.6 was $0.95 and $4.
Per task, the picture flips. Artificial Analysis measures K3 at $0.94 per Intelligence Index task, close to GPT-5.6 Sol's $1.04 and roughly half of Opus 4.8's $1.80, though far above GLM-5.2 ($0.32) and DeepSeek V4 Pro ($0.04).
|
|
cost per index task (Artificial Analysis)
|
| DeepSeek V4 Pro |
$0.04 |
| GLM-5.2 |
$0.32 |
| Kimi K3 |
$0.94 |
| GPT-5.6 Sol |
$1.04 |
| Claude Opus 4.8 |
$1.80 |
|
| 130M output tokens across the index vs a 63M average: about 2x the verbosity |
Source: Artificial Analysis, Intelligence Index v4.1, July 2026
|
|
|
The catch is how that $0.94 gets earned. K3 generated about 130 million output tokens across the index, roughly double the 63 million average for comparable models, at 62 tokens per second against a median near 71, with 1.99 seconds to first token. Running the full index cost $2,709.75 by Artificial Analysis's accounting. You cannot dial the thinking down yet, though Moonshot says more effort levels are coming.
One more number from the same report deserves a spot in your risk column. On AA-Omniscience, K3's accuracy rose from K2.6's 33 percent to 46 percent, but its hallucination rate also rose, from 39 percent to 51 percent. That reads like a model tuned to answer rather than abstain, which matters if your agent's output feeds anything downstream without review.
And the open part is still pending. The checkpoint is scheduled to go public on July 27, per the launch details. Until then there is no self-hosting and no second provider, and when the weights land, remember what they are: an MXFP4 model with 896-expert routing that Moonshot recommends serving on 64-plus accelerator pods. This is a cluster project, not a laptop pull.
|
|
Key insight: price the tokens, not the request. At $15 per million output and twice the median verbosity, a K3 agent loop spends Sonnet money, not open-model money, until prefix caching at $0.30 starts absorbing your repeated context.
|
|
|
Do this before July 27
Add a K3 row to your eval suite this week, through Moonshot's OpenAI-compatible endpoint or OpenRouter. Pin temperature 1.0 and top-p 1.0 to match the launch table, run your longest agent task first, and log output tokens per task next to pass rate. You want your own verbosity number before the weights conversation starts. The runnable harness config I use for rows like this, with the exact sampling settings, lives in the paid archive.
|
|
|
quick hits
|
The pelican tax. Independent testers measured 13,241 reasoning tokens for a single pelican drawing, about 25 cents per image. Verbosity shows up on request one, not at scale. Simon Willison's writeup has the details.
Day-one access. K3 is live on Moonshot's API, kimi.com, the mobile apps, Kimi Work, Kimi Code, and OpenRouter, no Moonshot account needed on that last one. The endpoint speaks the OpenAI wire format, so most stacks point at it with a base-URL change.
The size race has one leader. K3 is about 2.8x its 1T predecessor K2.6, and well past DeepSeek V4 Pro at 1.6T. By Moonshot's telling, Kimi models have defined the open-model size ceiling for nine of the past twelve months.
The politics arrived fast. The release landed ahead of the World AI Conference in Shanghai and fed the Washington debate over whether US teams should build on Chinese open weights. Three years of GPU export restrictions have not kept Chinese labs away from the frontier.
|
|
|
the take
Capability is no longer the moat. Serving economics and harness engineering are. A fourth-place model with open weights does not commoditize the frontier until someone other than Moonshot can serve 2.8 trillion parameters at these prices, and nobody else can even try before July 27. July 16 was the press release; July 27 is the event.
Moonshot benchmarked Kimi K3 in its own harness and every rival in theirs. The scores are real. So is the asterisk.
|
|
the open question
Suppose a second provider serves K3 at half of Moonshot's rate in August. Does your team move its agents, or does harness lock-in with Claude Code and Codex keep you where you are? Hit reply with your read. I read every reply.
|
|
|
The open-weights gap to the frontier is down to one model generation, with the checkpoint due in nine days at a price that undercuts the two models above it. Whatever you are building, your eval suite should have a K3 row by August.
Next issue: the new KAUST and Schmidhuber survey on self-improving agentic systems, read cover to cover so you can skip straight to the parts that apply to your stack.
|
|
|
|
