MiMo v2.6 Flash campaign publication summary¶
This is derivative publishing copy. The dated finding and linked artifacts are authoritative.
Canonical facts¶
- Model identity:
XiaomiMiMo/MiMo-V2.6-Flash-RL@3b38d063180c3e4aed9691fdc735f3d10b266ee4. - Runtime identity: SGLang v0.5.20 /
94602c9c2b7cbdb8efd5c52802dac6a1c180089e; image digestsha256:2b6c8893bb821bc9ac3fb95bc5c7ab8bddf748cf288a0ad6ad60ee96c49610ee. - Local outcome: MiMo V3 loaded with
flashinfer_mxfp4, then failed two bounded correctness probes with zero visible content; no capacity or throughput metric is eligible. - Baseline result: GLM's 32K nominal C4/16-word final completed 100/100 eligible requests; p50 visible TTFT was 21.98 s and p50 E2E was 26.03 s; exact figures are in the derived native summary.
- Important caveat: MiMo correctness failed; final GLM restoration is verified. The SWE native official summary is 0/1 with infrastructure confounds; no coding pass/fail is attributed solely to GLM.
- Context boundary: 327,680 tokens were requested and advertised; the measured full-attention KV pool held 91,342 tokens. No long-context acceptance was run.
- Decision: retain GLM;
no-promotion. No route, alias, or client change is authorized.
Claim ledger¶
| Public claim | Conditions | Evidence |
|---|---|---|
| GLM final C4 cell is eligible | 100 requests, nominal 32K, C4, unique prompts within cell, 16 controlled words | native final · summary |
| MiMo parser failure | first SGLang v0.5.20 load | log · status |
| MiMo OOM is startup-only evidence | corrected TP2 attempt; before weights/KV | recovered transcript · provenance |
| MiMo V3 correctness failure | weight loading and transport health passed; temperature-zero and official sampling diagnostics | sampling diagnostic · capture |
| No promotion | MiMo correctness/capacity gates fail or remain missing; final restoration verified | finding · summary |
Artifact manifest · Evidence index. This copy was saved locally; no external post was sent.