Kimi K3 Technical Report (Moonshot AI, July 2026)

"Kimi K3: Open Frontier Intelligence" is the technical report that Moonshot AI's Kimi Team released with the full weights of Kimi K3 on July 27, 2026, distributed as a PDF in the model's GitHub repository and linked from its Hugging Face model card (Source: github.com; huggingface.co). It shares a title with the July 16 launch announcement (Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026)) and supersedes that post's benchmark table with a larger one. The report describes a 2.8T-parameter mixture-of-experts model with 104 billion activated parameters, native vision and a one-million-token context window.

Architecture and training

The report attributes the model's design to two mechanisms from earlier Moonshot papers, Kimi Delta Attention (a hybrid linear attention paired with gated multi-head latent attention) and Attention Residuals, together with Stable LatentMoE, which activates 16 of 896 routed experts per token. Moonshot states that these and revised training and data recipes give about a 2.5-fold improvement in overall scaling efficiency over Kimi K2. Other components described are a SiTU-GLU activation, Quantile Balancing for expert load, a native vision encoder, and a Per-Head Muon optimizer.

Pre-training data spans web text, code, mathematics and knowledge, plus a vision corpus of captions, interleaved documents, OCR, perception, video and visual-coding data; knowledge and mathematics corpora are rephrased following the Kimi K2 recipe. Post-training consists of supervised fine-tuning, reinforcement learning across general, agentic and coding domains at several reasoning-effort levels, multi-teacher on-policy distillation, and quantization-aware training for MXFP4 weights. Infrastructure sections cover kernels and context parallelism for Kimi Delta Attention, balanced expert-parallel training, one-million-token agentic reinforcement learning with persistent sandbox state, and serving.

Evaluation approach

Public benchmarks are grouped into reasoning and knowledge, coding, agentic, and vision (Table 2). Kimi K3 is run at maximum reasoning effort and temperature 1.0, with top-p 0.95 for single-step tasks and 1.0 for agentic tasks. Baselines are Claude Fable 5, GPT-5.6 Sol, Claude Opus 4.8, GPT-5.5 and GLM-5.2, all at maximum effort except GPT-5.5 at "xhigh". The report states that Claude Fable 5 results "include fallback behaviors" and GPT-5.6 Sol results "include potential cyberguards." Coding models run under Kimi Code, Claude Code or Codex. Several comparison scores are not Moonshot's measurements: the report and model card cite Artificial Analysis (GDPval-AA v2, AA-Briefcase, τ³-Banking, Harvey Lab-AA, SciCode, AA-LCR, CritPt), Vals AI, and the official leaderboards for DeepSWE, Agents' Last Exam, Toolathlon-Verified, JobBench and APEX-Agents, dated July 23 or 24, 2026. Moonshot notes departures from official settings, including H20 rather than H100 GPUs for PostTrainBench and SWE-Marathon, and a context-compaction strategy for BrowseComp.

Table 3 reports sixteen in-house benchmarks covering coding, general agent and conversational experience, with footnotes giving the number of tasks on which Claude Fable 5 fell back or refused and GPT-5.6 Sol or GPT-5.5 refused. Table 4 gives blind expert preferences between Kimi K3 and Claude Opus 4.8 on web-development prompts.

A cybersecurity section evaluates two tiers: vulnerability discovery with proof-of-concept development, and end-to-end exploit development on an in-house suite of 36 expert-verified tasks, with GLM-5.2 as the only baseline. Moonshot says Anthropic and OpenAI frontier models refuse such tasks, "making a comparable evaluation infeasible." The section cites the joint UK AI Security Institute and CAISI assessment (UK AISI / CAISI Preliminary Assessment of Kimi K3's Cyber Capabilities (July 2026)) as consistent with its own results. A third-party section (Table 5) cites the Artificial Analysis Intelligence Index, the Vals Index and Arena leaderboards as of July 23, 2026, and a cost section compares score with per-task cost on four suites.

Findings as stated by Moonshot

The abstract says Kimi K3's overall performance "still trails the most powerful proprietary models, namely Claude Fable 5 and GPT-5.6 Sol," while exceeding the other evaluated models. The report identifies research-level reasoning (Humanity's Last Exam, CritPt) as a remaining gap and lists agentic suites on which it reports the highest score. On the in-house suite it names orchestration, deep research and coding as strengths and agent behaviour, multi-agent routing and agentic vision as areas where the model trails.

On cybersecurity, Moonshot reports that about 70% of human-reviewed vulnerability findings were confirmed, including 16 previously unknown vulnerabilities across six projects, and that the model solved 14 of 36 exploit tasks against 8 for GLM-5.2, with most successes on user-space targets. It describes a "clear gap to human experts" on hardened targets and regards its evaluation as "a lower bound on capability."

Reported results

Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.

Provenance

The raw file is a text parse of the PDF, captured on September 19, 2026; the title page parses without a date, and the date used is that of the weights-and-report release. Multi-line table cells parse imperfectly, and the Hugging Face model card provides a clean copy of Table 2 with per-benchmark sourcing footnotes. The report contains no refusal, jailbreak or other safeguard evaluations of Kimi K3. Confidence is medium: the results are developer-reported or developer-selected third-party figures, and several differ from the launch announcement's table published eleven days earlier.

Relationships

• supports: Kimi K3 — primary technical reference
• supersedes: Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026) — for the benchmark table only; the announcement remains the launch record
• related: Kimi K2 Technical Report — predecessor report
• related: UK AISI / CAISI Preliminary Assessment of Kimi K3's Cyber Capabilities (July 2026) — third-party cyber assessment the report cites