Skip to content
CrucibleMark

Kimi K3

Kimi K3 has been Moonshot AI’s Frontier flagship since July 2026, with 2.8 trillion total parameters making it the largest Open Weights model in the world to date. The Stable-LatentMoE architecture activates only 16 billion parameters per token from 896 experts; the context window spans 1,000,000 tokens for long-horizon agentic workflows. Kimi Delta Attention and Attention Residuals are designed to significantly improve scaling efficiency over the predecessor model.

Moonshot AI Version k3 Commercial use permitted MoE 2800 B (16 B active) 1000 K Context 02/2026 $3 / $15 per 1M

  • Open Weights
  • Frontier
  • OpenRouter
  • Text
  • Vision
  • Video
  • Long Context
  • Agentic Orchestrator
  • Batch

Sovereign Risk: HIGH Moonshot AI is a Chinese company subject to China’s National Security Law (NSL), which may enable state access to data via services hosted in China. In February 2025, Germany’s BSI explicitly warned against the use of Chinese AI cloud services; this risk assessment applies here conservatively as well. Kimi K3 has been announced as a 2.8T Open Weights model whose weights are scheduled for release on July 27, 2026; the operational risk arises primarily from cloud usage under Chinese jurisdiction, not from the mere existence of open weights.

Tool-use profile: 6 assets in detail

Comparison of asset performance (P1/P2/Combined) against the fleet average

Asset performance (radar)

Score Breakdown vs. Fleet Average

Tool-use details

Asset performance, reliability, and runtime profile

Name P1 P2 Combined
EU License Research 100 60 76
HTTP Fetch & Extract 80 80 80
Tool Failure Handling (404) 80 80 80
Web Search & Tool Selection 95 80 88.3
URL Construction & Fetch 80 80 80
Multilingual Search & Synthesis 100 60 80

CrucibleMark evaluates tool use across 6 independent tests. Click a test name for details.

Reliability

  • Tool Call Valid: No
  • Retry: Not required
  • Hallucination: Not detected

Reliability measures how consistently a model actually executes tool calls: Tool Call Valid schema and format accepted, Retry Required successful only after retry, Hallucination Flag fabricated tools or parameters detected. All three green means production-ready.

Runtime profile

Call 1
7.39
First Request
MCP
1.19
Protocol Latency
Synthesis
14.91
Response Generation
Total
140.95
Sum of All Phases
Token
20225
Input + Output
Cost
$0
Cost per Run

The runtime profile shows the latency and cost metrics for the model run: Call 1 First Request, MCP Protocol Latency, Synthesis Response Generation, Total sum of all phases, supplemented by Token input and output and Cost cost per run.

Tool-use review

· Long Context · Agentic Orchestrator

Deployment Verdict

Conditional deploy, because tool execution is strong and no hallucination was detected, but the invalid tool call limits confidence in an unsupervised MCP pipeline despite an overall good result.

Tool Execution Profile

Kimi K3 demonstrates clear orchestration strength. Tool execution is solid at P1 88.33, but not clean enough for a zero-supervision release, because at least one tool call was invalid. The critical factor is tool selection: in the Web Search and Tool Selection test — which checks whether the model chooses search over fetch without a hint — the model makes the right decision with high confidence. This argues against rigid pattern behavior and in favor of genuine tool selection. In the URL Construction and Fetch test — which measures autonomous derivation of a target URL — it remains usable but less deterministic. That is precisely where the limit shows: K3 plans well, but last-mile call precision is not consistent enough for strictly schema- or URL-sensitive systems. On the positive side, no retry was required. This looks more like a localized protocol or precision issue than a fundamental comprehension deficit.

Synthesis Fidelity

How well does it condense tool results? Solid, but not strong enough for high-stakes knowledge pipelines. P2 73.33 shows usable summarization across all assets, with noticeably weaker condensation on EU License Research and Multilingual Search and Synthesis. K3 extracts and connects information adequately, but loses sharpness on compliance-adjacent or cross-lingual condensation.

Does it stay within tool results or fall back on training data? The trust signal here is better than the condensation score. In the EU License Research honeypot — which checks whether current license restrictions are answered from web sources rather than training knowledge — no hallucination was detected. The model therefore stays on the tool track in principle. The P2 60 reflects imprecise condensation rather than unauthorized fallback to prior knowledge.

Error Resilience

Acceptable for production. In the 404 test — which requires transparent error communication rather than fabricated replacement content — K3 does not hallucinate page content. P2 80 is the right signal here: the model communicates failures clearly enough and does not damage the pipeline through fabricated results.

Operational Profile

Total 140.95s per run. Call 1: 7.39s. MCP latency: 1.19s. Call 2: 14.91s. Slow for the performance delivered. Cost/run: local.

Conclusion & Recommendation

Suitable for agentic research and orchestration pipelines with human review or downstream validation, particularly where tool selection matters more than perfect final synthesis. Not the first choice for fully automated compliance, registry, or URL-sensitive workflows where every call must be protocol-compliant and every synthesis step exact. In cloud deployments, significant jurisdictional risk applies; for sensitive data, only a strictly controlled, self-hosted setup is justifiable.

This analysis was generated automatically based on benchmark data. Model used: GPT-5.4 by OpenAI. The raw data and full methodology are documented in the GitHub project.