Skip to content

MiniMax M2.7 REAP

Current status and review date

Decision snapshot

  • Product role: historical measured Heavy challenger retained for comparison; not a current recommendation or live role.
  • Selected or best-qualified configuration: community MiniMax M2.7 REAP NVFP4 checkpoint on NGC vLLM 0.19, FP8 KV, 65,536 tokens, concurrency one, and thinking disabled.
  • Measured hardware: one RTX PRO 6000 Blackwell Max-Q card.
  • Evidence: functional, capacity, and bounded quality, with a historical-invalid immutable-identity limitation; 97.2 tok/s and the campaign's only 2/2 intelligence match to the Heavy baseline.
  • Decision: no-promotion; it was the best measured Heavy candidate at base-round close and was superseded in the campaign extension.
  • Important limitation: exact checkpoint revision and immutable image digest were not retained; the community prune's provenance and long-tail quality remain unaudited.
  • Review dates: retained evidence through 2026-07-11; dossier-format review 2026-08-31.

Open the retained recipe registry or jump to the decision, known limitations, or dated evidence.

Review narrative

2026-07-10–11 — Bakeoff result and supersession

The community REAP checkpoint ran through NGC vLLM 0.19 on one RTX PRO 6000 with compressed-tensors NVFP4, FP8 KV, 65,536 context, one sequence, the minimax_m2 tool parser, and thinking disabled. It passed context, tools, session recall, and both deterministic intelligence checks, the only candidate in the base round to match the Heavy baseline's 2/2. The warm 8K/256-output probe measured 97.2 tok/s and 86 ms TTFT.

The lane used 94.3 GB at 64K and had no headroom for a 131K test. Default thinking leaked think-text because no reasoning parser was configured. The community checkpoint had no retained exact revision, and its provenance and long-tail quality were not audited. Nemotron Puzzle later superseded it in the campaign extension with official provenance, 131K evidence, and a stronger controlled long-generation result.

Outcome: retain the measured result as historical comparison evidence with no-promotion; it is not an exact-rerun or current recommendation claim.

Immutable identity

  • Model repository: dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4.
  • Served name: minimax-m27-reap-139b-nvfp4.
  • Checkpoint revision: Not retained. The snapshot is historical-invalid for exact reruns.
  • Runtime: NGC vLLM 0.19 / image reference nvcr.io/nvidia/vllm:26.04-py3.
  • Immutable image digest: Not retained.
  • Community provenance/license: Not recorded as audited evidence.

Tested hardware and topology

  • Measured: one RTX PRO 6000 Blackwell Max-Q card on Primary Node.
  • Execution mode: isolated single-card candidate endpoint, one admitted sequence.
  • Capacity boundary: 94.3 GB observed at the 65,536-token configuration; 131K was not tested because no safe headroom remained.

Engine, quantization, KV, context, and concurrency recipe

Historical single-card lane

  • Engine: NGC vLLM 0.19.
  • Weights and KV: compressed-tensors NVFP4 and FP8 KV.
  • Contract: 65,536 tokens, one sequence, minimax_m2 tool parser, thinking disabled.
  • Recipe: registry entry and retained experiment Compose.

Evidence by measurement class

Functional and bounded quality

  • Status: functional and bounded quality with historical-invalid immutable identity.
  • Measured: context, tools, and session checks passed; intelligence 2/2.
  • Evidence: campaign finding.

Capacity and performance

  • Status: bounded capacity at 65,536 tokens.
  • Measured: 64K retrieval passed with 14.3-second TTFT; the warm 8K lane measured 97.2 tok/s and 86 ms TTFT.
  • Limit: the result is not a 131K or controlled cross-model quality claim.
  • Evidence: campaign finding and evidence index.

Decision and promotion state

Retained

  • Historical result: no-promotion; retain as the base-round Heavy comparison and a record of the community checkpoint's measured behavior.

Superseded or incomplete

  • Campaign recommendation: superseded by Nemotron Puzzle in the extension.
  • Exact rerun: blocked by the missing checkpoint revision and image digest.

Failures and gotchas

Evidence and provenance limits

  • Immutable identity: the checkpoint revision and image digest were not retained.
  • Community checkpoint: prune provenance, licensing, and long-tail quality were not audited.

Runtime and capacity limits

  • VRAM ceiling: 94.3 GB at 64K left no safe 131K path on the measured card.
  • Reasoning parser: none was configured; default-thinking responses leaked think-text, so the measured contract kept thinking disabled.
  • Comparison boundary: 97.2 tok/s remained roughly half the dated production Heavy baseline and did not authorize promotion.

Dated run history