MiMo v2.6 Flash replacement qualification¶
Date: 2026-09-21. Decision: retain GLM, no-promotion.
Result card¶
The GLM control completed the bounded 32K/C4 final cell. MiMo V3 loaded with the memory-scale remedy but failed both bounded correctness probes with zero visible content.
| Setup | Retained value |
|---|---|
| Baseline | GLM-5.3-Flash EXL3 r10 APC, TP2/C4, nominal 32K unique-prompt cell |
| Candidate | MiMo v2.6 Flash RL, pinned SGLang TP2/C1 reconstruction |
| Hardware | Two RTX PRO 6000 Blackwell Max-Q cards, native Linux |
| Candidate runtime | SGLang v0.5.20; --moe-runner-backend flashinfer_mxfp4; official mixed MXFP4/FP8, no speculation |
| Candidate context | 327,680 requested/advertised; 91,342 full-attention KV tokens allocated, C1 |
| Measurement | Direct baseline and bounded MiMo negative correctness probes |
| Decision | Retain GLM; no promotion or route change |
| Headline measurement | Local result | Conditions |
|---|---|---|
| GLM final C4 | 100/100 eligible; p50 TTFT 21.98 s, E2E 26.03 s | nominal 32K, 16 controlled words, C4; short-answer service cell |
| MiMo valid functional responses | 0 | V3 weight loading and transport health passed; both probes exhausted 9,216 reasoning tokens with zero visible content |
Important caveat: the 128-word GLM scout failed strict output and is not pooled with the 16-word cells. The final C4 p99 is one N=100 order statistic, not a stable service-level tail. Final restoration is verified; SWE grading is complete but confounded.
Evidence index · Artifact manifest · publication summary.
Outcome and method¶
The initial SGLang flag was ambiguous. V2 then ended with CUDA OOM in AudioProjection. V3 changed only to SM120 flashinfer_mxfp4, reached health, and then failed both bounded correctness probes. The corruption cause is unknown: TP4-to-TP2 fused-QKV handling and manually selected FlashInfer are investigation surfaces only. Template/parsers match the card, explicit official sampling also failed, and all 48 QKV-scale tensors resolved on both ranks. This is configuration-specific evidence, not proof that every MiMo runtime or recipe is infeasible.
GLM's direct six-check preflight and the final 100-request C4 cell are retained baseline evidence. The final cell follows an eight-request scout with the same canary seed, so its first eight requests may be warm; engine cache-hit fields were unavailable. No MiMo performance bar or comparative graph is valid.
The official vLLM 208 GB reference is text-only deployment guidance, not a physical proof that all TP2 recipes fail. No hardware-matched externally validated MiMo v2.6 recipe was found.
Evidence boundary¶
The public bundle preserves the run plan, workload manifest, and friction log. MiMo V3 has bounded negative correctness evidence but no capacity, performance, vision, or coding result. The native SWE official summary is 0/1 with explicit infrastructure confounds. Final GLM restoration is verified, including exact identity/config-hash parity, direct six-check preflight, and authenticated routed smoke.