Skip to content

MiMo v2.6 Flash

2026-10-15 MOPD replacement lanes: r4 is a diagnostic historical-comparison lane and r5 did not clear the frozen coding gate. The completed campaign retained GLM; this adds no causal sampling, capacity, performance, route, or promotion claim. Finding

Current status and review date

Status: not-qualified, no-promotion. Measured: pinned 5711b268 on the LIL vLLM image passed direct preflight 25/25 and frozen SWE 4/5. Native agentic results were 15/18 at vendor sampling, 17/18 greedy, and 17/18 thinking-off, below the required 18/18 gate. Limits: configured 327,680 tokens, C4, and eight images were not measured as actual context, concurrent capacity, or multimodal acceptance. No performance result exists. Evidence: 2026-10-06 scout and sanitized raw artifacts. The raw greedy capture shows the model itself emitted one parallel tool call and the parser preserved it; this is not parser-loss evidence. Exact r11 restoration is verified: direct 25/25, routed 7/7, configuration hashes unchanged, and router readmission complete.

Decision snapshot

  • Product role: unqualified candidate.
  • Selected or best-qualified configuration: none; the latest pinned vLLM no-spec TP2/C4 scout passed protocol and coding checks but failed the agentic floor.
  • Measured hardware: two RTX PRO 6000 Blackwell Max-Q cards, native Linux.
  • Evidence: latest preflight 25/25 and frozen SWE 4/5; agentic 15/18, 17/18, and 17/18 across three tested request policies. The September SGLang failure remains historical evidence.
  • Decision: no-promotion; GLM retained.
  • Important limitation: no policy passed the required 18/18 agentic gate, so full performance, context, C4, modality, and endurance qualification did not run.
  • Review dates: latest vLLM scout 2026-10-06; historical SGLang trial 2026-09-21.

Review narrative

2026-09-21 — bounded V3 correctness failure

The memory-scale remedy made transport health and load pass for the TP2 candidate, but both bounded probes produced no visible answer. Capacity and throughput testing stopped because correctness did not pass. Outcome: incomplete candidate, no-promotion.

Immutable identity

  • Latest model: XiaomiMiMo/MiMo-V2.6-Flash-RL@5711b268169967567844e1e560e8a3966da959b1.
  • Latest runtime: LIL vLLM 23f2a1830f3d34e5587e501bb6c953bab8e7799f, image digest ad7b059336e539068fc8b829fee7e7db04b0c3d9c32c455e69056d704619c729; public reconstruction.
  • Historical model: XiaomiMiMo/MiMo-V2.6-Flash-RL@3b38d063180c3e4aed9691fdc735f3d10b266ee4.
  • Historical runtime: SGLang v0.5.20, commit 94602c9c2b7cbdb8efd5c52802dac6a1c180089e, pinned image in the public reconstruction.

Tested hardware and topology

  • Measured: two RTX PRO 6000 Blackwell Max-Q cards; TP2 bounded trial.
  • Comparability boundary: no valid request-level performance or capacity result because correctness failed.

Engine, quantization, KV, context, and concurrency recipe

The latest vLLM reconstruction specifies official mixed MXFP4/FP8, FP8 KV, TP2/DCP1, no speculation, 327,680 total tokens, C4 admission, and eight images. Those configured limits did not pass full workload qualification.

The historical SGLang reconstruction specifies official mixed MXFP4/FP8, TP2, 327,680 context and C1. Both reconstructions are public documentation, not executable operator recipes.

Evidence by measurement class

Quality scout, 2026-10-06

Preflight passed 25/25 and frozen SWE resolved 4/5. Agentic results were 15/18 at vendor sampling, 17/18 greedy, and 17/18 thinking-off, below the required 18/18 floor. Every policy remains not-qualified; no comparative performance or full context/capacity/modality result exists. Finding and native artifacts.

Startup compatibility

  • Status: compatibility-only and bounded negative correctness.
  • Measured: V3 transport health and load passed at 81.41 GiB/rank with flashinfer_mxfp4; both temperature-zero and official sampling probes exhausted 9,216 reasoning tokens with zero visible content.
  • Limits: 91,342 full-KV tokens at default SWA ratio 0.8, not requested 327,680; no capacity, performance, vision, or coding score.
  • Evidence: dated finding, sampling diagnostic, and recovered OOM transcript.

Decision and promotion state

Promotion remains human-gated

No promotion, route, or client change is authorized.

Rejected, superseded, or incomplete

  • vLLM no-spec TP2/C4: rejected on the agentic floor despite SWE 4/5; no promotion or speed claim.
  • SGLang TP2 v2: incomplete after startup OOM; a supported remedy requires new pinned evidence.

Failures and gotchas

  • Parser: legacy --cuda-graph-max-bs is ambiguous in SGLang v0.5.20.
  • OOM and correctness: AudioProjection was the V2 final failed allocation. V3 flashinfer_mxfp4 passed transport health and load, then failed correctness; the corrupt-output cause is unknown. Fused-QKV TP handling and FlashInfer are investigation surfaces, not proven causes.
  • vLLM prior: the 208 GB text-only reference is not a universal TP2 feasibility proof.

Dated run history