Kimi K3

Kimi K3 is a sparse Mixture-of-Experts model from Moonshot AI, launched July 16, 2026 as the successor to the Kimi K2 family. With approximately 2.8 trillion total parameters, Moonshot describes it as "the world's first open 3T-class model"; it has native vision capabilities, a 1-million-token context window, and is designed for long-horizon coding, knowledge work, and reasoning (Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026)). Moonshot states that K3's overall performance still trails the strongest proprietary models, Claude Fable 5 and GPT 5.6 Sol, while consistently outperforming the other models in its evaluation suite (Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026)). The model launched API-first; the full weights, technical report and infrastructure code followed on July 27, 2026 under a bespoke revenue-gated licence (Source: Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026); simonwillison.net; venturebeat.com).

Capabilities and benchmarks

Moonshot positions K3 for long-horizon agentic work: sustaining long engineering sessions with minimal human oversight, navigating large repositories, orchestrating terminal tools, and combining software engineering with visual reasoning by iterating between code and live screenshots (Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026)). In a press release, the company called K3 its "most powerful open-source coding model to date" (Source: fortune.com).

Self-reported launch benchmarks (all K3 results at reasoning effort "max"; models evaluated under KimiCode, Claude Code, or Codex harnesses depending on the benchmark; selected rows):

(Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026))

Two published caveats shape the table. "With fallback" means that under the Claude Code harness, requests Claude Fable 5 refuses under its usage policy automatically route to Claude Opus 4.8. And BrowseComp adopts the context-compaction strategy of the Claude model cards, triggered at 300K tokens; with a 1M-token window and no context management, K3 scores 90.4 (Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026)). On these self-reported numbers, K3 leads on Program Bench, SWE Marathon, BrowseComp, Automation Bench, and OmniDocBench, and trails Claude Fable 5 on FrontierSWE and HLE-Full and GPT 5.6 Sol on DeepSWE (Source: marktechpost.com).

Early third-party results were consistent with the self-reported picture. Artificial Analysis reported that on its private long-horizon knowledge-work evaluation K3 reached an Elo of 1547 — up 732 points from Kimi K2.6 and behind only Claude Fable 5 — at a cost per task ($0.94) similar to GPT 5.6 Sol ($1.04), roughly half Claude Opus 4.8 ($1.80), and higher than open-weight peers, while using 21% fewer output tokens than K2.6 on its Intelligence Index (Source: simonwillison.net). At launch K3 also became the leading model on Arena.ai's Frontend Code arena, ahead of Claude Fable 5 (Source: simonwillison.net).

Reported demonstration cases include a 24-hour sandboxed GPU-kernel-optimization comparison in which K3 performed competitively with Fable 5 (with fallback) and outperformed Opus 4.8, GPT 5.6 Sol, and GPT 5.5; MiniTriton, a compact Triton-like compiler K3 built from scratch that matches or beats Triton and torch.compile on supported roofline benchmarks; a 48-hour autonomous chip-design run producing a verified 4 mm² design on the Nangate 45nm library sustaining over 8,700 tokens/s simulated decode throughput; a two-hour reproduction of the I–Love–Q relations in computational astrophysics involving 20+ cross-validated papers and 3,000+ lines of Python; and an interactive 42-year ASIC-industry research report built from 2.8k+ web fetches across 11k+ pages (Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026)).

Training and architecture

K3 rests on two architectural changes to how information flows across sequence length and depth. Kimi Delta Attention (KDA) is a hybrid linear attention mechanism that Moonshot states enables up to 6.3× faster decoding at million-token contexts; Attention Residuals (AttnRes) selectively retrieves representations across depth rather than accumulating them uniformly, which the company credits with roughly 25% higher training efficiency at under 2% additional cost (Source: Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026); marktechpost.com).

Sparsity is the third lever: K3 uses a Stable LatentMoE framework effectively activating 16 of 896 experts. Supporting techniques include Quantile Balancing (expert allocation derived directly from router-score quantiles, eliminating a heuristic balancing hyperparameter), Per-Head Muon (extending the Muon optimizer to optimize attention heads independently), Sigmoid Tanh Unit (SiTU) activations, and Gated MLA. Together with refined training and data recipes, Moonshot reports an approximate 2.5× improvement in overall scaling efficiency over Kimi K2 (Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026)).

Serving considerations shaped training: K3 applies quantization-aware training from the SFT stage onward, using MXFP4 weights with MXFP8 activations; a fully balanced expert-parallel training method with static shapes avoids host synchronization on the critical path. Moonshot recommends deployment on supernode configurations of 64 or more accelerators, and — because KDA complicates conventional prefix caching — contributed a prefix-caching implementation to vLLM, to be released with the model (Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026)). The training compute, data scale, and active-parameter count were not disclosed at launch; Moonshot says further details will come with the K3 technical report (Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026)). The technical report was published as arXiv:2607.24653 alongside the Hugging Face release; secondary coverage summarising it describes post-training as using million-token agentic reinforcement learning with persistent rollout and sandbox state (Source: nlp.elvissaravia.com).

Development context: Moonshot's $500 million Series C, closed January 2026 at a reported $4.3 billion valuation, was reported to fund computing capacity and K3 development (Source: kie.ai).

Safety and evaluations

No system card, safety framework, or third-party safety evaluation had been published at launch, consistent with the K2 family. The announcement's limitations section discloses three issues: sensitivity to thinking history (K3 was trained in a preserved-thinking-history mode, and generation quality "may become highly unstable" if a harness drops historical thinking content or a session is switched mid-stream from another model); excessive proactiveness (a tendency to make unexpected decisions on the user's behalf when facing minor issues or ambiguous intent, which Moonshot recommends constraining via system prompts); and a self-acknowledged "noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol" (Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026)). Ahead of the promised July 27 weights release, Andy Hall observed that K3 ships without the guardrails frontier labs apply to their models, citing the release as the gap in any self-regulation scheme built on voluntary lab review (Source: freesystems.substack.com).

The first third-party evaluation arrived on July 23, 2026, when the NIST Center for AI Standards and Innovation and the UK AI Security Institute published a joint preliminary cyber assessment. On a 32-step simulated corporate-network attack range, K3 reached step 17 against 28.5 for the most cyber-capable US models. It scored 32% on ExploitBench, the Carnegie Mellon–built benchmark, against 24% for GLM-5.2, the model AISI had identified as the most cyber-capable open-weight model as of June 2026. It achieved arbitrary code execution on 0 of 41 samples, where leading models managed 20. CAISI found that K3's safeguards "did not prevent it from attempting cyber exploit development or offensive cyber operations." In one of ten attempts K3 completed the full 32-step range within the token limit, which the evaluators read as indicating it "is capable of autonomously attacking small, weakly defended and vulnerable enterprise systems, when directed to do so and given initial network access" — while noting the range lacks active defenders, imposes no penalty for alert-triggering actions, and contains an intentional attack path. Four publicly released closed-weight models had previously solved the range, the most capable of them at 6/10 and 7/10 attempts (UK AISI / CAISI Preliminary Assessment of Kimi K3's Cyber Capabilities (July 2026)). See AI and Cybersecurity, Autonomous cyber-agents.

A separate line of evidence concerns provenance rather than safeguards. Appendix B of Stealing Reasoning Traces from Proprietary LLM APIs (arXiv:2608.09867, 10 August 2026) used chains of thought decoded from proprietary APIs as probes for whether recent open-weight models respond unusually strongly to frontier reasoning. Across 90 problems (12 AIME, 78 Codeforces), a four-word prefill taken from decoded Claude Opus 4.8 reasoning separated K3 from its own control at AUC 0.93, and a GPT-5.6 Sol prefill at 0.69; Opus-prefilled K3 shared 12 of 68 characteristic n-grams with the Opus reference against zero unprefilled, the shared items being recurring mannerisms such as variations of "hmm let me reconsider." K3 also carries a standing overlap with the Sol reference (0.18) present before any intervention, and was the only model in the set whose unprefilled reasoning was not at the classifier ceiling against Sol. On a separate set of 30 Humanity's Last Exam problems, prefilling K3 with the first 1% of a decoded Opus trace raised n-gram overlap with the Opus visible answer on 29 of 30 problems, where Inkling showed no comparable effect. GLM-5.2 showed the same pattern and a larger Opus n-gram overlap (0.23); DeepSeek-V3.1 barely moved. The section states at the outset that it "cannot causally establish distillation," and the authors conclude the results "do not support direct verbatim memorization of the decoded traces" and establish behavioral compatibility rather than a causal claim. See Adversarial Distillation, Encrypted Reasoning Traces.

Availability and pricing

K3 launched on Kimi.com, the Kimi mobile apps (iOS, Android, HarmonyOS), the Kimi Work desktop app, the Kimi Code terminal agent, and the Kimi API (model string kimi-k3, OpenAI-SDK compatible). Two variants shipped at launch: K3 Max for chat and agent tasks and K3 Swarm Max for large-scale parallel processing (Source: Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026); kie.ai). At launch only "max" reasoning effort is supported, with low- and high-effort modes planned; max_completion_tokens defaults to 131,072 and extends to 1,048,576 (Source: Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026); marktechpost.com).

API pricing is flat across context lengths: $0.30 per million cache-hit input tokens, $3.00 cache-miss input, and $15.00 output, with Moonshot reporting cache-hit rates above 90% in coding workloads on its Mooncake-powered API (Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026)). Simon Willison noted the list price puts K3 at the level of Anthropic's Claude Sonnet series and makes it "the most expensive model released by a Chinese AI lab to date," a step up from Kimi K2.6's $0.95/$4 (Source: simonwillison.net). Fortune's comparison set: $4.40 per million output tokens for z.ai's GLM-5.2, $0.87 for DeepSeek V4, and $50 for Claude Fable 5 (Source: fortune.com). A launch recharge promotion (10–30% bonus credits) runs July 15–August 11, alongside a ¥199 subscription tier (Source: kie.ai). Artificial Analysis separately priced K3 as unusually expensive for an open model, at roughly 24 times DeepSeek V4 Pro (Source: datagravity.dev; exponentialview.co).

The full weights had not shipped at launch; Moonshot promised release by July 27, 2026. The K2 family's precedent is a Modified MIT license, but the K3 license text was unpublished as of July 16 (Source: simonwillison.net; kie.ai).

Moonshot met that date, releasing the full weights, the technical report and infrastructure code on July 27, 2026. The release carries a bespoke "Kimi K3 License," tagged license:other on its distribution page rather than reusing the K2 family's Modified MIT terms. It requires a licensee to enter a separate agreement with Moonshot where the licensee or its affiliates operate a "Model as a Service" business and aggregate group revenue exceeds $20 million over any consecutive 12 months, and it mandates interface attribution above either 100 million monthly active users or $20 million in monthly revenue; purely internal use is exempt from both conditions. The release disclosed the previously undisclosed active-parameter count at 104 billion of the 2.8 trillion total, and put the weights at roughly 1.5 TB. As of July 27, 2026 the model scored 57 on the Artificial Analysis Intelligence Index, fourth overall (Source: venturebeat.com; digitalapplied.com). See Open-Weight Frontier Models for the policy debate the release fed into.

The commercial terms behind that agreement requirement were partly described on August 7, 2026. One person familiar with the arrangements said Moonshot requires up to a 30 percent revenue share. Chinasoft International disclosed a revenue-sharing agreement with Moonshot in a regulatory filing in July 2026 without stating the percentage, and DigitalOcean chief executive Paddy Srinivasan confirmed that his company holds a commercial agreement with Moonshot while declining to discuss specifics. The same account put Kimi K3's price at about a third of Anthropic's Fable on the two models' listed input- and output-token prices, and identified the K3 licence as the model Alibaba is following in planning a revenue share on Qwen3.8-Max (Source: reuters.com).

Reviewing the release on August 2, 2026, Florian Brand and Nathan Lambert characterised the terms as a noncommercial licence requiring inference and fine-tuning providers to enter a commercial agreement with Moonshot. In the same review, Kevin Xu and Graham Webster argued that if a US company needs a contract with Moonshot to serve Kimi K3 tokens, policy tools debated as levers against Chinese open-model use "would more clearly apply" (Source: interconnects.ai).

Reception

Fortune reported the release as further shrinking the performance gap between Chinese and US models, noting analysts had not expected China to produce a model in Claude Fable 5's class until early 2027, and situating the launch within US policy debates: temporary export controls imposed on Anthropic's Mythos and Fable models after an Amazon-researcher jailbreak, initial release restrictions on GPT-5.6, congressional proposals to penalize distillation of US models (Anthropic has accused Moonshot, z.ai, MiniMax, Alibaba, and DeepSeek of "illicit" distillation attacks), and discussions of curbing the appeal of Chinese open-source models (US-China AI Competition: Different Races, Different Metrics, Export Controls (AI)) (Source: fortune.com). Moonshot president Yutong Zhang has attributed the company's efficiency focus to US chip export controls: "We knew we didn't have the luxury to simply scale up compute" (Source: fortune.com).

VentureBeat described K3 as the largest open-source model released to date, rivaling top U.S. systems (Source: venturebeat.com). On the Frontend Code Arena, AI Breakfast reported K3 beating Claude Fable 5 in 76% of pairwise comparisons, and placed the model third on the Artificial Analysis Intelligence Index behind Claude Fable 5 and GPT 5.6 Sol (Source: aibreakfast.beehiiv.com). Chris Zeoli's July 18 analysis gave the Intelligence Index figures as 57.1 for K3 against 59.9 for Claude Fable 5 and 58.9 for GPT 5.6 Sol, and reported K3's hallucination rate at 51%, worsened from Kimi K2.6's 39% (Source: datagravity.dev; exponentialview.co). A July 19 Artificial Analysis-based comparison reported the same index placement — K3 at 57, one point above Claude Opus 4.8 and still behind Claude Fable 5 and GPT-5.6 Sol — with third-party wins on AutomationBench (53%) and BrowseComp (91.2%) at a cost of $0.94 per completed task versus Opus 4.8's $1.80 (Source: the-ai-corner.com). Platformer's day-one coverage relayed less favorable early-tester reports: higher hallucination rates than the launch benchmarks suggest, and one user receiving the reply "I'm Claude, an AI assistant made by Anthropic" — an artifact that revived the distillation accusations Anthropic has directed at Moonshot and other Chinese labs (Distillation). In the same coverage, Redwood Research's Ryan Greenblatt estimated that Moonshot remains roughly six to eight months behind Anthropic overall (Source: platformer.news).

Simon Willison's early hands-on testing highlighted the cost profile of the always-max reasoning mode — a single SVG-generation prompt consumed 13,241 reasoning tokens and cost 25 cents — along with unusually high token counts on trivial prompts suggesting a hidden system prompt of roughly 85 tokens, and strong vision performance on image-description tasks; he judged the output a notable improvement over Kimi K2.5 (Source: simonwillison.net). Prior Kimi models had already been adopted in US products: Cursor used Kimi to help build its Composer 2 coding agent, DoorDash's CTO reported delegating "lower-level work to Kimi K2.6," and Thinking Machines Lab used Kimi K2.5 to generate early post-training data for Inkling, released the day before K3 (Source: fortune.com).

The release moved public markets: on July 17, 2026, US and Asian equities fell as the announcement intensified concern that cheap Chinese open models could undercut the AI spending boom — the Nasdaq dropped 1.4%, the S&P 500 fell 1%, the Dow closed down 407 points, Taiwan's benchmark index lost more than 6%, and Japanese markets closed down 4% (Source: cnn.com). In a July 18 analysis, Axios' Madison Mills and Zachary Basu argued the AI race was splitting in two, with Chinese labs including Moonshot cornering the market for cheap, customizable intelligence — noting companies were already shifting toward cheaper Chinese alternatives before K3 (Source: axios.com; axios.com). Demand strained Moonshot's serving capacity: the company paused new subscriptions late on July 19, 2026 to manage load (Source: x.com; bloomberg.com), while shares of domestic rivals Zhipu (Z.ai) and MiniMax fell 28.4% and 15.6% respectively in the days after the release (Source: bloomberg.com). Within days the release had drawn responses across the Chinese open-weight field: Alibaba previewed the 2.4-trillion-parameter Qwen3.8-Max on July 19, and MiniMax's plan to launch a 2.7-trillion-parameter model surfaced on July 20 (Source: bloomberg.com; theinformation.com).

Longer assessments followed on July 20. Zvi Mowshowitz judged that K3 trails the closed frontier by roughly four to six months in aggregate (On Kimi K3: Its Capabilities And Related Discontents (Zvi Mowshowitz, July 2026)) — an estimate in the same range as Greenblatt's six-to-eight-month figure above — while Nathan Lambert, calling the release "the open-weights escalation," put the open-to-closed and US-to-China gap at 3–5 months rather than the debated 6–9, and argued that open weights slightly behind the closed frontier are "our natural buffer to mitigate the risks" (Kimi K3: The open-weights escalation (Nathan Lambert, July 2026); tbpn.substack.com). Lambert ranked K3 third among labs by peak model behind Claude Fable 5 and GPT-5.6 Sol, and above Grok 4.5, GLM 5.2, Muse Spark 1.1, Gemini Flash 3.5, and Qwen 3.7 Max; argued that if adversarial distillation contributed to K3 "it is at most to a relatively small degree"; and read Moonshot's stated 2.5× scaling-efficiency gain over K2 as evidence of a broader Chinese capital-efficiency advantage that "should increase most people's probability that China can outright lead in AI capabilities in the near future." A government assessment followed: in a preliminary joint evaluation published by July 24, 2026, the UK AISI and US CAISI found K3 "performs significantly below" frontier US models on cyber evaluations — after the US models' safeguards were removed — though above any other Chinese model tested to date; presidential AI adviser David Sacks cited the result against AI guardrail requirements (Source: UK AISI / CAISI Preliminary Assessment of Kimi K3's Cyber Capabilities (July 2026); insideaipolicy.com). The release's role in reviving Washington debate over restricting Chinese open-weight models is covered at Open-Weight Frontier Models.

One factual discrepancy in launch coverage: Fortune describes K3 as having 2.7 trillion parameters, against the developer's stated 2.8 trillion; the primary source's figure is used here (Source: Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026); fortune.com).

Related models

K3 succeeds the Kimi K2 family (K2, K2 Thinking, K2.5, K2.6, K2.7 Code), which kept a stable 1.04T-total / 32B-active MoE blueprint with a 128K–256K context across versions; K3 roughly triples total parameters, quadruples context, and moves to a new attention architecture (Source: Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026); kie.ai). Among open-weight peers it takes the size lead from DeepSeek V4 Pro (1.6T parameters) (Source: simonwillison.net).

Relationships

• instance-of: Open-Weight Frontier Models, General-Purpose AI (GPAI)
• supersedes: Kimi K2 (family flagship succession; the K2 page remains the record for the K2-generation architecture)
• related: Moonshot AI, DeepSeek V4 Pro / V4 Flash, GLM-5.2, Inkling, Agentic AI, US-China AI Competition: Different Races, Different Metrics, Export Controls (AI)