Kimi K3: Open Weights Arrive at the Frontier

Kimi K3 is Moonshot AI’s new flagship: 2.8 trillion parameters, a 1-million-token context window, native multimodality, and a set of vendor-reported benchmarks that put it ahead of Claude Opus 4.8 and GPT-5.5 on most agentic suites — and within a point or two of Claude Fable 5 and GPT-5.6 Sol on several more. It launched July 17, 2026 across Kimi.com, Kimi Work, Kimi Code, and the Kimi API, with open weights promised by July 27.

That last clause is the story. Closed labs have shipped stronger models this summer, and open-weight labs have shipped cheaper ones. What hasn’t happened before is a model this close to the closed frontier arriving with a dated commitment to publish its weights — Moonshot bills K3 as the first 3-trillion-parameter-class model to make that promise. If the July 27 release lands as described, the gap between what you can rent from a closed vendor and what you can run under your own control narrows to months, not years.

This post walks through what shipped, what’s genuinely new in the architecture, what the benchmark charts do and don’t establish, the pricing and developer mechanics, and — the part we care most about for clients — what an open-weights-committed near-frontier model means for anyone currently negotiating with exactly two frontier vendors.

The announcement came via Moonshot’s official channels on July 17, 2026, alongside a technical blog and updated platform documentation. The launch surfaces span the consumer product (Kimi.com), the workplace tier (Kimi Work), the terminal coding agent (Kimi Code), and the developer API — all live on day one, which is itself a departure from the staggered rollouts closed vendors have favored this year.

K3 arrives five weeks after K2.7-Code, Moonshot’s June coding release, and it consolidates a lineage that has been iterating in public all year — K2.5’s agent swarms in January, K2.6 in April, K2.7-Code in June. The K2 line established Moonshot’s pattern of shipping weights under a permissive license after a short commercial exclusivity window; K3 formalizes that pattern with a date printed in the launch post. Worth being precise, though: as of publication the weights are not yet released, and the license terms for K3 specifically are unconfirmed. Prior Kimi releases used a Modified MIT license; whether K3 follows suit is a July 27 question.

Thinking is always on and runs at maximum effort by default — Moonshot says lower effort modes are coming later. That default matters for reading everything that follows: the benchmark charts label all models as maxed out on thinking effort, and the pricing math in section 04 assumes reasoning tokens are part of every response.

Strip the branding and K3’s technical pitch is about how information flows through a very large, very sparse model — across sequence length, and across depth. Two named mechanisms carry most of the load, per Moonshot’s technical blog.

Kimi Delta Attention (KDA) targets the long-context serving problem. Moonshot claims up to 6.3x faster decoding in million-token contexts — the regime where conventional attention makes long-context models technically available but economically unattractive. If the claim holds up in third-party serving benchmarks, it is the difference between a 1M window you quote in marketing and one your agents actually use on every turn.

Attention Residuals (AttnRes) works the depth axis: residual pathways for attention state that Moonshot credits with roughly 25% higher training efficiency at under 2% additional cost. Combined with the sparsity jump — 16 of 896 experts under a Stable LatentMoE framework — and refined training recipes, Moonshot puts the overall scaling-efficiency gain at about 2.5x over K2. These are vendor engineering claims, not audited figures, but they are specific, falsifiable ones — the kind the open-weights release will let outsiders actually test.

The deployment story is equally concrete: weights ship in MXFP4 with MXFP8 activations for hardware compatibility, and Moonshot recommends supernode configurations of 64 or more accelerators for serving. Hold that number for section 05 — it does a lot of work in the self-hosting conversation.

Every figure below is Moonshot-reported at launch, with all models run at maximum thinking effort (max or xhigh) per the chart footnotes. No independent replication exists yet — for a model whose weights open in ten days, it will come quickly, which is itself a reason vendor charts from open-weight labs tend to be better behaved than closed ones. With that framing, the shape of the results is consistent: K3 clears Opus 4.8 and GPT-5.5 on most agentic suites, beats everything on a handful, and trails Fable 5 and GPT-5.6 Sol on the hardest software-engineering and knowledge-work evals.

Three honest readings of that table. First, the wins are real wins on the vendor’s own terms — Program Bench, SWE Marathon, BrowseComp, SpreadsheetBench 2, and Automation Bench are led by K3 on charts that include the two strongest closed models available. Second, the losses cluster where it matters most for hard engineering work: Fable 5’s FrontierSWE lead is 5.4 points and its GDPval-AA Elo lead is 92 points — not rounding error. If you want the deeper context on how those two closed flagships compare with each other, our frontier comparison covers that matchup. Third, Moonshot also reports internal knowledge-work benchmarks (Online Exp Bench 75.5, DECK-Bench 73.5, Finance-Bench 62.6) where K3 at max effort outperforms Opus 4.8 and GPT-5.5 — but internal benches from any vendor deserve the heaviest discount, and we’d treat those as directional only.

The API pricing is where the architecture claims become a commercial argument: $3.00 per million input tokens on cache misses, $0.30 on cache hits, and $15.00 per million output tokens — flat across the full 1M-token context, with no long-context surcharge or tiering. Moonshot cites cache-hit rates above 90% in coding workloads, which, if representative of your traffic, pulls the effective input price toward the $0.30 figure. Compare the standing rates: Fable 5 at $10/$50, GPT-5.6 Sol at $5/$30, and GPT-5.5’s 1M window carrying a long-context surcharge above 272K input tokens that K3’s flat pricing simply doesn’t have.

The usual caveat applies with extra force here: price per token is not price per task. K3 thinks at maximum effort on every request at launch — reasoning_effort currently accepts only max, with low and high modes promised later — so its token consumption per job will run high relative to models that modulate effort. Until third parties publish tokens-per-task comparisons, treat the bars above as list prices, not landed costs.

For developers, the surface is deliberately boring in the good sense. The API is OpenAI-SDK compatible (model ID kimi-k3), supports strict JSON-schema structured output, tool calling with required-choice and dynamic tool loading, streaming with separate reasoning and content deltas, and a partial mode that continues generation from a text prefix. Output defaults to 131,072 tokens and can be configured up to 1,048,576. Sampling is fixed — temperature 1.0 and top-p 0.95, non-negotiable — which removes a whole class of knob-tuning from evaluation plans, whether you like that or not.

Be clear-eyed about what the open-weights commitment does and does not mean. It does not mean you’ll run K3 on a workstation: at 2.8T parameters — even in MXFP4 — Moonshot’s own serving recommendation is a supernode of 64 or more accelerators. This is not a self-host-on-your-own-GPU story in the way smaller open-weight coding models are. For nearly every organization, “running K3 yourself” will mean renting managed capacity from an inference provider that hosts the open weights — the same consumption model DeepSeek V4 established this spring.

What the weights release does buy is structural, and it accrues even to companies that never download a single shard. Hosted-price competition: within days of a major open-weights drop, multiple providers historically list the model and undercut the first-party API. Continuity insurance: a model you can obtain from many hosts — or, at enterprise scale, hold in escrow yourself — cannot be suspended, repriced, or deprecated by a single vendor decision. And auditability: the architecture claims in section 02 stop being marketing the day outsiders can measure them. The license is the open question — Modified MIT would match Moonshot’s K2 precedent, but until July 27 the terms are unconfirmed, and any commercial-use plan should wait for the actual text.

Our standing advice on open-weight challengers hasn’t changed with this release — it has strengthened. We laid out the full argument in our open-weight second-source playbook: you don’t adopt a challenger model because its launch chart beat your incumbent by a point; you qualify it as a second source so that pricing, access, and deprecation decisions by your primary vendor stop being existential. K3 is the strongest candidate yet for that role — near-frontier vendor scores, a public pricing sheet at under a third of Fable 5’s output rate, an OpenAI-compatible API that makes evaluation nearly free to wire up, and a dated open-weights commitment that removes the single-vendor failure mode entirely.

The honest counterweights: benchmarks are vendor-reported for at least another week or two; max-only thinking effort inflates per-task cost in ways the price list hides; Moonshot itself concedes the product-polish gap versus Fable 5 and GPT-5.6 Sol; and for regulated workloads, a Beijing-based provider’s hosted API raises data-governance questions that only the self-hosted weights — under a confirmed license, on infrastructure you control — actually resolve. None of those are reasons to ignore the release. All of them are reasons to evaluate it on your own tasks rather than adopt it from a chart.

For the businesses we work with, the practical takeaway is about posture. AI capability is compounding on both sides of the open/closed line, and the buyers getting the best terms are the ones with credible alternatives wired up before renewal conversations start. That’s an architecture decision as much as a procurement one — routing, fallbacks, eval harnesses, and cost governance that treat any single model as replaceable. Designing that posture is the core of our AI transformation practice, and it’s how we run our own content engine across providers — this release goes into our candidate pool the same way we’re recommending it goes into yours.