Skip to content
CrucibleMark

Xiaomi MiMo V2.6 Flash

The smaller sibling of the MiMo-V2.6 series prioritizes efficiency over maximum size: MiMo-V2.6-Flash-RL by Xiaomi activates around 15 billion of 309 billion parameters per token, matching the flagship Pro on agent and coding tasks at roughly one-third of the API price. Omnimodal for text, image, video, and audio, context up to 1 million tokens, Open Weights under the MIT license.

Xiaomi Version V2.6-Flash Commercial use permitted MoE 309 B (15 B active) 1024 K Context $0.14 / $0.28 per 1M

  • Open Weights
  • Server
  • OpenRouter
  • Text
  • Vision
  • Video
  • Audio
  • Agentic Orchestrator
  • Long Context
  • Interactive

Sovereign Risk: MEDIUM Xiaomi releases MiMo-V2.6-Flash-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 (including the National Intelligence Law and the Data Security Law), which remains relevant when using the model via Xiaomi’s own API platform; with purely local self-hosting, the operational risk is substantially reduced.[434][445]

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 20 52
HTTP Fetch & Extract 80 35 57.5
Tool Failure Handling (404) 80 80 80
Web Search & Tool Selection 100 80 91
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: 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
4.7
First Request
MCP
0.98
Protocol Latency
Synthesis
19.93
Response Generation
Total
153.65
Sum of All Phases
Token
14277
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 · Agentic Orchestrator · Long Context

Deployment Verdict

Conditional deploy, because tool execution is strong, but an invalid tool call and weak synthesis fidelity limit confidence in production tool pipelines. The overall impression is usable, but not cleared for unsupervised high-trust workloads.

Tool Execution Profile

MiMo V2.6 Flash demonstrates genuine tool intelligence. On the Web Search & Tool Selection test — which checks whether the model distinguishes between search and direct retrieval without being prompted — it selects the correct tool with confidence. This argues against rigid pattern-matching behavior. Execution-side stability also holds on the Multilingual Search & Synthesis test.

Precision at the final step is weaker. On the URL Construction test, which checks correct derivation of a target URL followed by a fetch, it reaches only a solid but non-deterministic level. On top of that, at least one tool call overall was not valid. This is not a retry problem and therefore not mere format noise — it is a protocol risk: in MCP pipelines, the call must be structurally correct on the first attempt.

Synthesis Fidelity

How well does it condense tool results? Only to a limited degree. P2 performance at 59.17 is the clear bottleneck. EU License Research and HTTP Fetch & Extract in particular show that the model does not reliably convert retrieved content into sound, precise statements. In production systems, what matters is not only that a tool was used, but that the response correctly and completely reflects the tool’s findings.

Does it stay within the tool result, or does it fall back on training? On the Honeypot EU License Research test — which probes exactly this trust failure — it does not hallucinate overtly, but the result remains clearly too weak at P2=20. This is a security risk, not merely a quality issue: when a model fails to cleanly bind current web findings, it can produce statements in compliance or policy pipelines that appear well-founded but are in fact unverified.

Error Resilience

On the 404 test, MiMo V2.6 Flash responds in a production-ready manner. It communicates the failure transparently and does not fabricate page content. This behavior is precisely what robust orchestration requires, because the pipeline can then handle the error in a controlled way.

Sovereignty Profile

Locally deployable and operationally attractive for sovereign deployments. At 73.42, it sits 0.00 points above the fleet average of 68.17. The MIT-licensed open weights are a genuine advantage, especially since provenance risk is substantially reduced in self-hosted setups.

Conclusion & Recommendation

Suitable for agentic pipelines where tool selection, web research, and transparent error handling matter more than precise final synthesis. Not suitable for compliance, license review, fact-critical extraction, or other workloads where the model must summarize tool findings exactly and in an auditable manner. Recommendation: deploy as an orchestrator or research front-end with downstream verification — not as the final trust-bearing synthesis layer.

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.