{"slug": "kimi-k3-a-forensic-analysis", "title": "Kimi K3: A Forensic Analysis", "summary": "Moonshot AI announced Kimi K3 on 2026-07-16, a 2.8-trillion-parameter mixture-of-experts model with 896 total experts and 16 active per token, achieving an Artificial Analysis Intelligence Index score of 57 (#4 overall) and a #1 Frontend Code Arena Elo of 1679, the first first-place arena finish for a Chinese-developed model. Open weights and the technical report are due 2026-07-27 on Hugging Face; pricing is set at $3.00 per million input tokens and $15.00 per million output tokens, with cache hits at $0.30 per million tokens.", "body_md": "# Kimi K3: A Forensic Analysis\n\nA claim-by-claim audit of Kimi K3: architecture, benchmarks, pricing, and hardware.\n\nMoonshot AI announced Kimi K3 on 2026-07-16. Open weights and the technical report are due 2026-07-27. Neither is published yet, so architecture figures below are vendor-stated. Each claim in this audit is marked **Verified** (independent measurement), **Vendor-stated** (awaiting inspection), or **Not disclosed**.\n\n## Release facts\n\n| Item | Value | Status |\n|---|---|---|\n| Announcement | 2026-07-16 | Verified |\n| Weights + technical report | Due 2026-07-27 on Hugging Face | Pending |\n| Access today | API only (Moonshot platform, OpenRouter) | Verified |\n| Variants | K3 Max; K3 Swarm Max (~300 sub-agents, ~4,000 steps) | Vendor-stated |\n| Input price | $3.00 / M tokens; $0.30 cache hit | Verified |\n| Output price | $15.00 / M tokens, reasoning tokens included | Verified |\n| License | Unpublished | Not disclosed |\n\n## Architecture\n\n| Spec | Value |\n|---|---|\n| Total parameters | 2.8T |\n| Active parameters | Not stated; ~50B community estimate |\n| Experts | 896 total, 16 active per token (1.8%) |\n| Shared experts | Not stated |\n| Routing | LatentMoE, Quantile Balancing |\n| Attention | KDA (linear) + gated MLA (full); layer ratio not stated |\n| Residual scheme | Attention Residuals: cross-layer retrieval |\n| Context | 1M tokens |\n| Vision | Native input |\n| Optimizer | Per-Head Muon |\n| Precision | MXFP4 weights, MXFP8 activations, QAT from SFT stage |\n| Checkpoint size | ~1.4TB at MXFP4 |\n| Hidden size, layers, vocab, tokenizer, positional encoding | Not disclosed |\n| Training tokens | Not disclosed; ~2.5x scaling efficiency vs K2 claimed |\n\n*Figure 1. Vendor-described block structure. Most blocks use KDA (linear attention); a subset uses gated MLA (full attention). Depth and interleave ratio are not disclosed; the predecessor architecture Kimi-Linear used 3:1.*\n\n*Figure 2. Vendor-described LatentMoE routing. Expert scoring runs in a projected latent space. 16 of 896 experts fire per token (1.8%). Shared-expert count is not stated.*\n\nVendor-reported effects: KDA up to 6.3x faster decoding at 1M context; Attention Residuals ~25% training efficiency gain at <2% overhead.\n\n## Lineage\n\n| K2 (Jul 2025) | K2.5 (Jan 2026) | K3 (Jul 2026) | |\n|---|---|---|---|\n| Total parameters | 1T | 1T (1.1T with vision) | 2.8T |\n| Active per token | 32B | ~32B | Not stated (~50B est.) |\n| Experts | 8 of 384 + 1 shared | 8 of 384 | 16 of 896 |\n| Attention | MLA | MLA | KDA + gated MLA |\n| Context | 128K | 256K | 1M |\n| Vision | None | Native | Native |\n| Optimizer | MuonClip | MuonClip lineage | Per-Head Muon |\n| Native precision | BF16 | INT4 QAT (Thinking) | MXFP4 QAT |\n\n*Figure 3. K-series release timeline. The final entry is a commitment, not a release.*\n\n## Benchmarks\n\n| Measurement | Vendor | Independent | Source |\n|---|---|---|---|\n| Artificial Analysis Intelligence Index | n/a | 57 (#4 overall) | Artificial Analysis |\n| Terminal-Bench 2.1 | 88.3 | 84.6 | Artificial Analysis |\n| Frontend Code Arena | n/a | #1, 1679 Elo | Arena.ai |\n| Output speed | 260 tok/s | 62 tok/s | Artificial Analysis |\n| Cost per completed task | n/a | $0.94 | Artificial Analysis |\n\n- Index 57: highest recorded for an open-weight model. Level with Opus 4.8. Below Claude Fable 5 and GPT-5.6 Sol. Above GLM-5.2 (51) and DeepSeek V4 Pro (44).\n- Frontend Code Arena #1: first first-place arena finish for a Chinese-developed model.\n- Vendor table mixes four harnesses: Kimi Code, Claude Code, Codex, mini-SWE-agent. Rows are not directly comparable.\n- Absent from vendor table: SWE-bench Verified, AIME, tau-bench.\n- SWE-bench figures in circulation: 76.8% and 93.4% (Vals AI). Different harnesses. No same-harness reproduction published.\n- Launch figures are upper bounds until reproduced on the target endpoint and harness.\n\n*Figure 4. Independent aggregate scores. K3 leads all open-weight models; closed frontier models remain ahead. Opus 4.8 value is approximate as published. Exact values are in the table above.*\n\n### Hallucination vs K2.6 (AA-Omniscience)\n\n| Metric | K2.6 | K3 |\n|---|---|---|\n| Accuracy | 33% | 46% |\n| Hallucination rate | 39% | 51% |\n\nMore correct answers, and more incorrect assertions in place of declining. Moonshot documentation notes “overly proactive” ambiguity resolution. Test against your workload before unattended agentic use.\n\n## Cost\n\n| Item | Value |\n|---|---|\n| Measured verbosity | ~2x median output tokens (Artificial Analysis) |\n| Full AA evaluation cost | $2,690.80 |\n| Single SVG generation (Willison) | $0.25; 13,241 of 16,658 output tokens were reasoning |\n| Reasoning budget exhaustion | 4 of 20 OpenRouter completions returned no answer |\n| Per-task comparison | K3 $0.94; GPT-5.6 Sol $1.04; Opus 4.8 $1.80; DeepSeek V4 Pro $0.04 |\n| Reasoning effort tiers | Max at launch; standard and high added after |\n\nNew consumer subscriptions are suspended. Moonshot cites GPU constraints.\n\n## Hardware\n\n| Precision | Checkpoint size |\n|---|---|\n| MXFP4 (native) | ~1.4TB |\n| BF16 | 5.6TB |\n| 2-bit quantization (estimate) | 650GB to 1TB |\n\n- Minimum deployment: ~64-accelerator scale-up domain (SemiAnalysis estimate). Exceeds one DGX B200 node at 4-bit.\n- No GGUF or MLX conversions exist. No weights to convert.\n- KDA and LatentMoE are not in stable vLLM or SGLang. A KDA vLLM implementation is reported to land with the weights.\n- Local deployment: not feasible at this model size.\n\n## Harness contract\n\n- K3 is trained with reasoning traces preserved across turns.\n- Dropping or compacting thinking history degrades output quality (Moonshot statement).\n- Officially supported harnesses: Kimi Code, Claude Code, OpenCode, Roo Code.\n- Same weights produce different results under different context management. Background:\n[the agentic harness](/blog/the-agentic-harness-architecture/).\n\n*Figure 5. The harness contract. K3 expects prior reasoning traces in context on later turns. The degraded path is Moonshot’s own characterization.*\n\n## Claim ledger\n\n| Claim | Status |\n|---|---|\n| 2.8T total, 16-of-896 MoE, KDA, 1M context | Vendor-stated |\n| Active parameter count | Not disclosed |\n| Highest open-weight score on an independent index | Verified |\n| #1 Frontend Code Arena | Verified |\n| Overall parity with closed frontier | Not shown; independent aggregates place K3 below Fable 5 and GPT-5.6 Sol |\n| 260 tok/s | Not reproduced; 62 tok/s measured |\n| Hallucination change vs K2.6 | Verified: 39% to 51% |\n| Open weights and license | Unpublished; due 2026-07-27 |\n| Local deployment feasibility | Not feasible at stated size |\n\n## Verification plan for the weights release\n\n*Figure 6. Audit state machine. Every vendor-stated claim moves to Verified or Discrepancy once the weights and report are inspectable.*\n\n| Item | Test |\n|---|---|\n| Expert counts, attention layout, undisclosed dimensions | config.json vs launch blog |\n| License terms | Published text vs K2 modified MIT |\n| Checkpoint size | ~1.4TB expected at MXFP4 |\n| Launch benchmarks | Fixed-harness reproduction |\n| Training data provenance | Technical report data section |\n\nProvenance context: Anthropic’s February 2026 report states Moonshot harvested 3.4M+ Claude exchanges via fraudulent accounts. No public evidence connects that dataset to K3 training. Moonshot has not commented.\n\nResults of this checklist will be published here.\n\n## Scope\n\n- No evidence found that any Moonshot figure is wrong.\n- No first-hand capability testing performed yet.\n- This is an audit of public claims against public evidence, following the method from\n[our JEPA validation](/blog/what-we-learned-validating-jepa/).\n\nCurrent independent standings: [AI model matrix](/research/ai-models/).\n\n*Veso AI evaluates AI systems on evidence. Assessing K3 for your stack? Get in touch.*", "url": "https://wpnews.pro/news/kimi-k3-a-forensic-analysis", "canonical_source": "https://veso.ai/blog/kimi-k3-forensic-analysis/", "published_at": "2026-07-19 14:00:00+00:00", "updated_at": "2026-07-20 13:45:48.418585+00:00", "lang": "en", "topics": ["large-language-models", "ai-products", "ai-research", "artificial-intelligence"], "entities": ["Moonshot AI", "Kimi K3", "Hugging Face", "OpenRouter", "Artificial Analysis", "Arena.ai", "Vals AI"], "alternates": {"html": "https://wpnews.pro/news/kimi-k3-a-forensic-analysis", "markdown": "https://wpnews.pro/news/kimi-k3-a-forensic-analysis.md", "text": "https://wpnews.pro/news/kimi-k3-a-forensic-analysis.txt", "jsonld": "https://wpnews.pro/news/kimi-k3-a-forensic-analysis.jsonld"}}