GPT-OSS 120B¶

Current status and review date¶

Decision snapshot

Product role: historical capacity and throughput control; no-promotion, with no claim about current live routing.
Selected or best-qualified configuration: observed checkpoint revision with vLLM 0.23.1, native MXFP4, Marlin MoE, FP8 KV, and 131,072 configured tokens.
Measured hardware: one NVIDIA RTX PRO 6000 Blackwell Max-Q on Primary Node.
Evidence: functional preflight, 128K retrieval, tools 20/20, and a historical 183.2 tok/s controlled long-generation result.
Decision: retain as a historical throughput control only; no-promotion.
Important limitation: the contemporaneous 183.2 tok/s artifact does not retain a comparison-grade sequence cap, and the later deterministic quality recheck is historical-invalid because reasoning budgets were not comparable.
Review dates: retained evidence through 2026-07-12; dossier-format review 2026-08-31.

Review narrative¶

2026-07-10 — baseline throughput and capacity¶

The baseline passed the functional, long-context, tool, session, and bounded intelligence checks at a 131,072-token window. Its established controlled long-generation result was 183.2 tok/s, making it useful as a historical throughput reference rather than a current deployment recommendation.

2026-07-12 — deterministic protocol invalidation¶

The deterministic suite first produced empty visible answers because native GPT-OSS reasoning consumed the configured response budget. Raising only that budget changed the score, making the cross-model quality comparison invalid. The retained result therefore documents an evaluation-protocol failure, not a zero-quality model result.

Immutable identity¶

openai/gpt-oss-120b, observed revision b5c939de8f754692c1647ca79fbf85e8c1e70f8a. The tested vLLM image resolved to sha256:907377dddef392f6b679d9c071e1c33c3935b4dc993b61d0352e391a5319ff3e and reported engine version 0.23.1rc1.dev531+ga65f93fb2.

The oldest launch recipe did not pin --revision. Treat the observed revision as retained identity evidence, not proof that a future mutable pull is equal.

Tested hardware and topology¶

One RTX PRO 6000 Blackwell Max-Q on Primary Node as the earlier Heavy baseline, under Windows 11, Docker Desktop, and WSL2.

Engine, quantization, KV, context, and concurrency recipe¶

Historical measured shape¶

The measured service used vLLM 0.23.1, native MXFP4, Marlin MoE, FP8 KV, CUDA graphs, the OpenAI tool parser, and a 131,072-token served context. The comparison-grade maximum sequence setting for the 183.2 tok/s run was not retained.

Pinned reconstruction recipe¶

The current registry entry in configs/serve-recipes.toml pins the observed checkpoint revision, exact image digest, eight sequences, 8,192 batched tokens, and the retained serving flags. It is a reproducible recipe pinned to the observed identity; do not retroactively label every field in it as proven input to the earlier 183.2 tok/s artifact.

Evidence by measurement class¶

Functional and capacity¶

functional and capacity evidence includes coding, structured JSON, a 128K needle, and 20/20 shared-prefix tool calls. The July 12 conventional run completed 10/10 sequential 8K requests at c1.

Performance¶

The retained controlled long-generation result is 183.2 tok/s. A separate short mixed-prompt run measured 29.87 aggregate output tok/s, but that value is not a controlled decode rate.

Historical-invalid quality comparison¶

The deterministic recheck is not a valid protocol-v3 quality comparison. Native reasoning exhausted the original visible-answer budget, and raising only the response cap changed the result. Classify those cross-model scores as historical-invalid, not as model quality evidence.

Decision and promotion state¶

Retain GPT-OSS 120B as a historical throughput and compatibility control; no-promotion. The pinned reconstruction recipe supports controlled reuse but does not convert the invalid deterministic quality run into qualification.

Failures and gotchas¶

Historical launch identity¶

The oldest service did not pin --revision, and its 183.2 tok/s artifact did not retain a comparison-grade maximum sequence cap. The observed revision and later pinned recipe narrow reuse but do not erase those provenance gaps.

Reasoning-budget confounder¶

GPT-OSS ignores the Qwen-style enable_thinking=false control and continues to use its native reasoning channel. The deterministic suite's original token budget could end before visible content appeared. A valid cross-model protocol-v3 quality result was not recorded.