A claim-by-claim audit of Kimi K3: architecture, benchmarks, pricing, and hardware.
Moonshot 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.
Release facts #
| Item | Value | Status |
|---|---|---|
| Announcement | 2026-07-16 | Verified |
| Weights + technical report | Due 2026-07-27 on Hugging Face | Pending |
| Access today | API only (Moonshot platform, OpenRouter) | Verified |
| Variants | K3 Max; K3 Swarm Max (~300 sub-agents, ~4,000 steps) | Vendor-stated |
| Input price | $3.00 / M tokens; $0.30 cache hit | Verified |
| Output price | $15.00 / M tokens, reasoning tokens included | Verified |
| License | Unpublished | Not disclosed |
Architecture #
| Spec | Value |
|---|---|
| Total parameters | 2.8T |
| Active parameters | Not stated; ~50B community estimate |
| Experts | 896 total, 16 active per token (1.8%) |
| Shared experts | Not stated |
| Routing | LatentMoE, Quantile Balancing |
| Attention | KDA (linear) + gated MLA (full); layer ratio not stated |
| Residual scheme | Attention Residuals: cross-layer retrieval |
| Context | 1M tokens |
| Vision | Native input |
| Optimizer | Per-Head Muon |
| Precision | MXFP4 weights, MXFP8 activations, QAT from SFT stage |
| Checkpoint size | ~1.4TB at MXFP4 |
| Hidden size, layers, vocab, tokenizer, positional encoding | Not disclosed |
| Training tokens | Not disclosed; ~2.5x scaling efficiency vs K2 claimed |
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.
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.
Vendor-reported effects: KDA up to 6.3x faster decoding at 1M context; Attention Residuals ~25% training efficiency gain at <2% overhead.
Lineage #
| K2 (Jul 2025) | K2.5 (Jan 2026) | K3 (Jul 2026) | |
|---|---|---|---|
| Total parameters | 1T | 1T (1.1T with vision) | 2.8T | | Active per token | 32B | ~32B | Not stated (~50B est.) | | Experts | 8 of 384 + 1 shared | 8 of 384 | 16 of 896 | | Attention | MLA | MLA | KDA + gated MLA | | Context | 128K | 256K | 1M | | Vision | None | Native | Native | | Optimizer | MuonClip | MuonClip lineage | Per-Head Muon | | Native precision | BF16 | INT4 QAT (Thinking) | MXFP4 QAT |
Figure 3. K-series release timeline. The final entry is a commitment, not a release.
Benchmarks #
| Measurement | Vendor | Independent | Source |
|---|---|---|---|
| Artificial Analysis Intelligence Index | n/a | 57 (#4 overall) | Artificial Analysis |
| Terminal-Bench 2.1 | 88.3 | 84.6 | Artificial Analysis |
| Frontend Code Arena | n/a | #1, 1679 Elo | Arena.ai |
| Output speed | 260 tok/s | 62 tok/s | Artificial Analysis |
| Cost per completed task | n/a | $0.94 | Artificial Analysis |
- 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).
- Frontend Code Arena #1: first first-place arena finish for a Chinese-developed model.
- Vendor table mixes four harnesses: Kimi Code, Claude Code, Codex, mini-SWE-agent. Rows are not directly comparable.
- Absent from vendor table: SWE-bench Verified, AIME, tau-bench.
- SWE-bench figures in circulation: 76.8% and 93.4% (Vals AI). Different harnesses. No same-harness reproduction published.
- Launch figures are upper bounds until reproduced on the target endpoint and harness.
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.
Hallucination vs K2.6 (AA-Omniscience)
| Metric | K2.6 | K3 |
|---|---|---|
| Accuracy | 33% | 46% |
| Hallucination rate | 39% | 51% |
More 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.
Cost #
| Item | Value |
|---|---|
| Measured verbosity | ~2x median output tokens (Artificial Analysis) |
| Full AA evaluation cost | $2,690.80 |
| Single SVG generation (Willison) | $0.25; 13,241 of 16,658 output tokens were reasoning |
| Reasoning budget exhaustion | 4 of 20 OpenRouter completions returned no answer |
| Per-task comparison | K3 $0.94; GPT-5.6 Sol $1.04; Opus 4.8 $1.80; DeepSeek V4 Pro $0.04 |
| Reasoning effort tiers | Max at launch; standard and high added after |
New consumer subscriptions are suspended. Moonshot cites GPU constraints.
Hardware #
| Precision | Checkpoint size |
|---|---|
| MXFP4 (native) | ~1.4TB |
| BF16 | 5.6TB |
| 2-bit quantization (estimate) | 650GB to 1TB |
- Minimum deployment: ~64-accelerator scale-up domain (SemiAnalysis estimate). Exceeds one DGX B200 node at 4-bit.
- No GGUF or MLX conversions exist. No weights to convert.
- KDA and LatentMoE are not in stable vLLM or SGLang. A KDA vLLM implementation is reported to land with the weights.
- Local deployment: not feasible at this model size.
Harness contract #
- K3 is trained with reasoning traces preserved across turns.
- Dropping or compacting thinking history degrades output quality (Moonshot statement).
- Officially supported harnesses: Kimi Code, Claude Code, OpenCode, Roo Code.
- Same weights produce different results under different context management. Background:
the agentic harness. Figure 5. The harness contract. K3 expects prior reasoning traces in context on later turns. The degraded path is Moonshot’s own characterization.
Claim ledger #
| Claim | Status |
|---|---|
| 2.8T total, 16-of-896 MoE, KDA, 1M context | Vendor-stated |
| Active parameter count | Not disclosed |
| Highest open-weight score on an independent index | Verified |
| #1 Frontend Code Arena | Verified |
| Overall parity with closed frontier | Not shown; independent aggregates place K3 below Fable 5 and GPT-5.6 Sol |
| 260 tok/s | Not reproduced; 62 tok/s measured |
| Hallucination change vs K2.6 | Verified: 39% to 51% |
| Open weights and license | Unpublished; due 2026-07-27 |
| Local deployment feasibility | Not feasible at stated size |
Verification plan for the weights release #
Figure 6. Audit state machine. Every vendor-stated claim moves to Verified or Discrepancy once the weights and report are inspectable.
| Item | Test |
|---|---|
| Expert counts, attention layout, undisclosed dimensions | config.json vs launch blog |
| License terms | Published text vs K2 modified MIT |
| Checkpoint size | ~1.4TB expected at MXFP4 |
| Launch benchmarks | Fixed-harness reproduction |
| Training data provenance | Technical report data section |
Provenance 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.
Results of this checklist will be published here.
Scope #
- No evidence found that any Moonshot figure is wrong.
- No first-hand capability testing performed yet.
- This is an audit of public claims against public evidence, following the method from
[our JEPA validation](/blog/what-we-learned-validating-jepa/).
Current independent standings: [AI model matrix](/research/ai-models/).
Veso AI evaluates AI systems on evidence. Assessing K3 for your stack? Get in touch.