Launch update

Kimi K3 officially launches: open weights promised by July 27, early benchmarks put it near the frontier

Post-launch update on Kimi K3: the 16-of-896-experts architecture, the July 27, 2026 open-weights promise, and independent results from Artificial Analysis, Arena, and Vals AI.

17 Jul 2026 14 min read

Kimi K3 Moonshot Benchmark Open Weights

Kimi K3 changed fast within two days. On July 16, 2026 TechCrunch was still summarizing a Financial Times report that this 2-to-3-trillion-parameter model would arrive "in the coming days". A day later its API documentation went live, and shortly after that Moonshot published a technical blog, announced the launch on its official X account, and promised full model weights no later than July 27, 2026. Independent evaluators like Artificial Analysis, Arena, and Vals AI have also posted early results.

This article is an update to the earlier launch analysis. The rule stays the same: product facts, vendor claims, independent evaluator results, and promises not yet delivered are kept separate. The most precise status for the model weights as of July 17, 2026: open weights have been officially announced, but not yet released.

Specs and architecture: the MoE details are finally out

The API docs and the technical blog both describe K3 as a 2.8-trillion-parameter model with a context window of about one million tokens, native visual understanding, Kimi Delta Attention (hybrid linear attention), and Attention Residuals. Its targets are long-horizon coding, knowledge work, and deep reasoning.

The most important new information sits in the Mixture of Experts configuration. The technical blog states that K3's Stable LatentMoE activates 16 out of 896 experts for a given token. That number describes sparsity, not the active parameter count; the size of each expert and the non-MoE components are not detailed enough to compute it, and Moonshot has indeed not announced an active parameter count.

Moonshot also mentions quantization-aware training with MXFP4 weights and MXFP8 activations, plus techniques like Quantile Balancing and Per-Head Muon. The vendor claim: roughly a 2.5x improvement in overall scaling efficiency compared to Kimi K2. The full technical report is still pending, so outside parties cannot verify that efficiency claim yet.

API contract and pricing: unchanged since day one

This part is still identical to launch. The model ID is kimi-k3, the base URL is OpenAI-SDK compatible, thinking mode is always on, and reasoning_effort on the public API only supports max for now. Sampling parameters like temperature and top_p are fixed and cannot be changed. Output limits: max_completion_tokens defaults to 131,072 and can be raised to 1,048,576. Context caching runs automatically for stable prefixes.

Bar chart: Kimi API pricing comparison per 1M tokens — Cache hit, cache-miss input, and output prices per 1M tokens for K2.6, K2.7 Code, and K3.

Model	Cache hit	Input (cache miss)	Output	Context
Kimi K3 (`kimi-k3`)	$0.30	$3.00	$15.00	1,048,576
Kimi K2.7 Code	$0.19	$0.95	$4.00	256k
Kimi K2.6	$0.16	$0.95	$4.00	256k

Bar chart: Kimi context window comparison — K3's context window reaches 1 million tokens; K2.6 and K2.7 Code sit at 256k.

K3 offers roughly four times the context of K2.6 or K2.7 Code, at much higher input and output rates. The technical blog adds one operational claim: the official API reportedly reaches a cache hit rate above 90% on coding workloads. That is a vendor metric that depends heavily on workload shape and prefix stability, so treat it as an indication, not a guarantee.

There is one reproducibility note developers should know. The public API docs state top_p is fixed at 0.95, while the benchmark footnote in the technical blog says K3 results were run with temperature 1.0 and top-p 1.0. Until Moonshot explains the difference, the vendor benchmark table should not be treated as exactly reproducible through regular API requests.

Open weights: officially promised for July 27, not available today

This is the biggest status change since the first analysis. Moonshot now calls K3 an open 3T-class model and, both in the technical blog and the launch post on X, commits to releasing full model weights no later than July 27, 2026. The phrase "open weights by 27 July" is an announced deadline, not proof the artifact exists.

As of July 17, no K3 repo has appeared on the moonshotai Hugging Face org or the MoonshotAI GitHub org. There are no weight files, no license, no model card, no deployment guide. Artificial Analysis, Arena, and Vals also still label K3 as proprietary, which is fair because they can only test the hosted API.

Even if the weights ship on schedule, self-hosting will not be trivial. Moonshot itself recommends deployment on a supernode with 64 accelerators or more, and Reuters notes that K3's scale makes self-hosting unrealistic for many organizations. For most developers the API will likely stay the main route; open weights matter more for research labs, cloud providers, and companies with large inference clusters.

Vendor benchmarks: strong at coding and agentic work, not a sweep

The technical blog publishes a wide benchmark table. Every number below is vendor-reported: harnesses differ between KimiCode, Claude Code, and Codex, so the comparison is not an identical bake-off. A sample:

Benchmark	Kimi K3 max	Claude Fable 5	GPT-5.6 Sol	Claude Opus 4.8
Program Bench	77.8	76.8	77.6	71.9
Terminal-Bench 2.1	88.3	84.6	88.8	84.6
FrontierSWE	81.2	86.6	71.3	66.7
SWE Marathon	42.0	35.0	39.0	40.0
BrowseComp	91.2	88.0	90.4	84.3
GPQA-Diamond	93.5	92.6	94.1	91.0
HLE-Full	43.5	53.3	44.5	49.8
OmniDocBench	91.1	89.8	85.8	87.9

This table does not support a headline like "K3 beats every American model". K3 leads on SWE Marathon, BrowseComp, and OmniDocBench, but trails badly on HLE-Full and loses to Claude Fable 5 on FrontierSWE. Moonshot itself admits K3's overall performance still sits below the strongest proprietary models in their own aggregate evaluation. That admission is rare in launch material, and it makes the rest of the table more worth reading.

Independent evaluators: one model, three stories

Artificial Analysis gives K3 a score of 57 on its Intelligence Index, ranked #4 of 189 models on its page when verified on July 17. Its performance profile is neither cheap nor fast: output speed around 62 tokens per second (below the comparison median), time to first token of 1.99 seconds (better than median), and the model is verbose. Running the Intelligence Index produced about 130 million output tokens at a cost of roughly $2,690.80, about twice the median. For agentic workloads that means price per token is not the only factor; measure cost per completed task.

Arena tells two different stories. On WebDev Overall, K3 ranks #1 with a score of 1679 (+17/-17), still preliminary, with about 1,757 votes as of July 16. On Text Arena Overall, K3 sits around #9 with a score of 1486 ±11 and about 3,024 votes. Both are snapshots that can shift as votes come in, and both should be read together: users strongly prefer K3's frontend output, but for general chat it is frontier-class without dominating.

Vals AI places K3 at #2 of 38 models with a Vals Index of 74.70% in its July 16 update. The components are consistent with the coding story: 95.10% on the SWE-bench Verified subset, 91.27% on the Vibe Code Bench subset, 80.90% on Terminal-Bench 2.1 (three full trials). Finance results are more moderate: 72.61% on CorpFin v2 and 55.88% on the Finance Agent v2 subset. Note the word subset; several Vals evaluations use subsets to keep testing fast, so the numbers are not always equivalent to full benchmarks.

How builders should read K3

Need	Conclusion as of July 17, 2026
Try Moonshot's strongest model via API	Available now: `kimi-k3` with `reasoning_effort: max`
Frontend / web prototypes	Strong external signal (Arena WebDev #1), but still preliminary
Agentic coding and terminal tasks	Strong on Vals and the vendor table
General chat	Frontier-class, not #1 on Arena Text
Self-host today	Not possible; weights and license not out yet
Prepare to self-host after July 27	Plan for a large cluster (vendor recommends 64+ accelerators)
Cost-sensitive workloads	Consider K2.x or hybrid routing; K3 is verbose

Closing

The next ten days are the interesting part. Kimi K3 now has a much fuller evidence package: official specs, the 16-of-896-experts architecture details, a vendor benchmark table, and early results from three independent evaluators. The honest picture is not "K3 wins at everything" but "K3 has entered the frontier group with standout strength in coding and agentic work", while trailing on some reasoning evaluations and running fairly slow and token-hungry.

One thing has not changed since the first analysis: the weights are not out. What changed is the status, from "unconfirmed" to an official commitment with a date. Until the files, license, and deployment guide actually ship, K3 remains a hosted model. Read the quickstart and pricing before running long workloads, test your own repositories with clear acceptance criteria, and re-check Moonshot's Hugging Face and GitHub on July 27, 2026.