Skip to content
CrucibleMark

Xiaomi MiMo V2.6 Pro

With 1.02 trillion total parameters, MiMo-V2.6-Pro-RL by Xiaomi is currently the largest model with fully open weights. Approximately 42 billion parameters activate per token, trained in a single mixed reinforcement learning run across coding, agent, and safety tasks; the accompanying live dashboard of the training process provided unusual transparency. Omnimodal for text, image, video, and audio, context up to 1 million tokens, MIT license.

Xiaomi Version V2.6-Pro Commercial use permitted MoE 1020 B (42 B active) 1024 K Context $0.435 / $0.87 per 1M

  • Open Weights
  • Frontier
  • OpenRouter
  • Text
  • Vision
  • Video
  • Audio
  • Instruction-Tuned
  • Agentic Orchestrator
  • Batch

Sovereign Risk: MEDIUM Xiaomi releases MiMo-V2.6-Pro-RL under MIT with fully open weights, which significantly improves operational provenance for local deployment. As a Chinese developer, however, Xiaomi remains subject to national laws, which remains relevant when using the model via Xiaomi’s own API platform; with local self-hosting, the operational risk is substantially reduced.[448][444]

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 40 64
HTTP Fetch & Extract 80 60 70
Tool Failure Handling (404) 80 35 64.3
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: 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
5.81
First Request
MCP
0.93
Protocol Latency
Synthesis
40.89
Response Generation
Total
285.77
Sum of All Phases
Token
14044
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

Created · Instruction-Tuned · Agentic Orchestrator

Deployment Verdict

Conditional deploy, because tool usage is strong but the model hallucinates on tool errors, undermining trust in an MCP pipeline. The overall impression is good, but the safety finding outweighs the combined score of 75.04.

Tool Execution Profile

Xiaomi MiMo V2.6 Pro demonstrates genuine tool intelligence rather than a rigid fetch pattern. In the Web Search & Tool Selection test — which checks the choice between search and direct retrieval without an explicit hint — it correctly identifies the need for web_search and achieves P1 100. This speaks to workable planning logic in dynamic pipelines. In the URL Construction test, which measures the derivation of a target URL from internal knowledge and the subsequent fetch, it performs solidly but not deterministically enough for high-strictness flows. P1 80 is good, but not hard-reliable. The critical protocol signal: Tool-Call valid is false. This does not mean it fundamentally misses tools, but the MCP output was not clean enough for smooth production operation in at least one relevant case. On the positive side, no retry was required. This looks more like a validity or format deficiency in the call than a deeper comprehension problem.

Synthesis Fidelity

How well does it consolidate tool results? Only moderately. P2 59.17 shows that MiMo often retrieves results correctly but fails to keep them precise enough during consolidation. This is visible in EU License Research with P2 40 and Tool Failure Handling (404) with P2 35. It performs better in Multilingual Search & Synthesis with P2 80 and in URL Construction & Fetch with P2 80.

Does it stay within the tool result or fall back on training data? In the honeypot EU License Research, it formally stays on the tool path and does not hallucinate there. This matters for compliance-adjacent research. Nevertheless, hallucination detected is globally set to true. This is not merely a quality deficiency — it is a security risk: when a model presents fabricated statements as tool output, the entire infrastructure becomes unreliable.

Error Resilience

The model clearly fails here. In the 404 test, which distinguishes transparent error communication from fabricated replacement content, MiMo hallucinates page content despite a failed tool call. This is production-critical without exception. An agent must only report the error state or propose a new search after a 404. It must not reconstruct content.

Operational Profile

Total 285.77s. Call 1 5.81s, MCP latency 0.93s, Call 2 40.89s. Slow for the performance shown. Cost/run: local. Price per model card is low to moderate for Frontier class, but the runtime erodes that advantage in interactive pipelines.

Conclusion & Recommendation

Suitable for supervised research and orchestration pipelines in which a controller validates tool responses, catches error states, and verifies final outputs. Not suitable for autonomous retrieval, compliance, or incident workflows where tool failures must be handled robustly and truthfully. If you deploy MiMo, do so only with strict response validation, hard gates on 4xx/5xx, and a policy that consistently discards unsupported synthesis.

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.