Skip to content
CrucibleMark

Kimi K2.6

Kimi K2.6 is Moonshot AI’s multimodal model for agentic tasks, coding, and tool-assisted workflows, with native input support for text, image, and video. The MoE architecture activates only 32 billion of the total one trillion parameters per token; the context window spans 256,000 tokens. Available as an Open Weights model locally or via cloud API, with Chinese jurisdiction as a material cloud risk factor.

Moonshot AI Version k2.6 Commercial use permitted MoE 1000 B (32 B active) 256 K Context 12/2025 $0.95 / $4 per 1M

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

Sovereign Risk: HIGH Moonshot AI is a Chinese company and subject to China’s National Security Law (NSL), which may enable state access to data. In February 2025, the BSI explicitly warned against the use of Chinese AI cloud services; this risk assessment conservatively applies here as well.

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 70 82
HTTP Fetch & Extract 80 80 80
Tool Failure Handling (404) 80 80 80
Web Search & Tool Selection 100 60 82
URL Construction & Fetch 80 80 80
Multilingual Search & Synthesis 100 80 90

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
9.55
First Request
MCP
1.19
Protocol Latency
Synthesis
44.05
Response Generation
Total
328.69
Sum of All Phases
Token
21481
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

Updated · Agentic Orchestrator · Long Context

Deployment Verdict

Conditional deploy: Kimi K2.6 is generally viable for tool-assisted pipelines because it does not hallucinate and performs strongly in tool execution, but inconsistently valid tool calls and only moderate synthesis fidelity limit confidence for strictly deterministic workflows.

Tool Execution Profile

The model demonstrates genuine tool selection rather than mere schema imitation. In the Web Search & Tool Selection test — which requires distinguishing between search and direct retrieval without an explicit hint — it correctly identifies the need for web_search. This speaks to usable orchestration logic in open MCP pipelines. In the URL Construction test, which checks whether the model derives a target URL from its own knowledge and subsequently retrieves it correctly, it remains usable but not precise enough for fully deterministic fetch chains.

The critical issue is less the selection than the protocol cleanliness. The global finding “Tool-Call valid: false” indicates that at least one call was not formally or structurally robust enough. Since no retry was required, this does not appear to be a fundamental comprehension problem. It is more an indication of edge-case imprecision in call form or parameterization. For production MCP environments, this means: plan for guardrails and tool wrappers.

Synthesis Fidelity

How well does it condense tool results? Only adequately, not strongly. The P2 performance shows a model that mostly draws researched content together in a usable way, but does not work with consistently sufficient precision for high-quality result synthesis. This aligns with the individual scores: solid extraction and good multilingual capability, but weaker condensation precisely where source selection and weighting matter.

Does it stay within the tool result or fall back on training data? Here the confidence signal is positive. In the EU License Research test — which checks whether current license restrictions are answered from web sources rather than training knowledge — Kimi K2.6 does not hallucinate. For compliance-adjacent research, this matters more than stylistic quality. It demonstrates that the model fundamentally respects the tool infrastructure.

Error Resilience

In the 404 test, which measures how the model handles a failing tool call, it responds transparently and does not fabricate page content. That is exactly what is acceptable in production. A failed retrieval thus remains visible as an error, rather than being silently rewritten into false facts.

Operational Profile

Call 1: 9.55s. Call 2: 44.05s. MCP latency: 1.19s. Total per run: 328.69s. Slow. Cost/run: local. Price: $0.95 per 1M input and $4.0 per 1M output. Not expensive for the performance shown, but clearly not a low-latency model.

Conclusion & Recommendation

Suitable for agentic research pipelines, multilingual tool chains, and workflows where transparent error handling matters more than perfect result synthesis. Not the first choice for strictly validated MCP pipelines with a hard dependency on formally clean tool calls, or for pipelines where the final synthesis is passed downstream without subsequent review. Deploy with schema validation, output checks, and clear tool ownership of facts.

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.