GPT-OSS Puzzle 88B¶

Current status and review date¶

Decision snapshot

Product role: historical human-approved Heavy profile and pinned rollback in its dated campaign; retained as reproducible evidence, not a claim about the current live rollback chain.
Selected or best-qualified configuration: exact Puzzle checkpoint and custom Anvil vLLM commit, Marlin MXFP4 MoE, FP8 KV, 131,072 context, eight sequences, and 8,192 batched tokens.
Measured hardware: one NVIDIA RTX PRO 6000 Blackwell Max-Q on Primary Node.
Evidence: full functional and router qualification, long-context retrieval, tools 20/20, capacity 40/40 at c8, and repeated quality with one retained unified-diff failure.
Decision: retain as a dated rollback configuration and historical Heavy compatibility result; no present-day route or promotion claim.
Important limitation: strict unified-diff formatting passed only 2/3, and the earlier 65.72% GPQA result was not rerun on the final image.
Review dates: retained evidence through 2026-07-18; dossier-format review 2026-08-31.

Review narrative¶

2026-07-17 — custom-engine qualification¶

The exact Puzzle checkpoint required a custom vLLM branch for heterogeneous expert counts, attention windows, and the missing Harmony <|call|> EOS behavior. The pinned image passed model identity, long-context retrieval, tools, Responses, bounded capacity, and the initial quality checks on one RTX PRO 6000.

2026-07-18 — final-image Heavy enablement¶

The final source-built image passed the managed functional, router, interface, capacity, and repeated quality gates and became the human-approved Heavy profile in its dated campaign. The strict quality suite retained one 2/3 unified-diff failure. That campaign later treated the profile as a pinned rollback; neither decision describes current live state.

Immutable identity¶

Model: nvidia/gpt-oss-puzzle-88B revision 9c0e0746a0d2218b28cc7b2cb3ce4e1a2f50fdb2.
Engine: fakoli/anvil-vllm commit 485463b3498ed3ffcf0c8fcb52c1670a21be5d82.
Qualified image ID: sha256:470f7b7e39c4363696d5a79fd041d6a45253229a9ba1c055d089ddbdc0ed120c.
Served name: gpt-oss-puzzle-88b.

Tested hardware and topology¶

One RTX PRO 6000 Blackwell Max-Q on Primary Node through an isolated serve and then managed Heavy/router qualification. The dated transaction preserved the Gemma 4 rollback configuration; it is not a current live-topology claim.

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

Pinned serving shape¶

The qualified profile uses the native Puzzle architecture, Harmony template, OpenAI tool parser, automatic tool choice, Marlin MXFP4 MoE, FP8 KV, 131,072 context, eight sequences, and 8,192 batched tokens. Preserve the explicit EOS override and V2-runner disable.

Complete reproduction procedure¶

The canonical operator recipe retains the exact engine build, checkpoint pull, image verification, GPU selection, managed start, functional/capacity/quality gates, router verification, and rollback procedure. Do not substitute stock vLLM or silently retag another build.

Evidence by measurement class¶

Functional and long-context¶

functional, capacity, and quality evidence covers coding, JSON, long-context retrieval, tools 20/20, streaming, Responses, session behavior, and timeout triage. The exact final image also closed the earlier tool-parser HTTP 500.

Capacity and latency¶

The production-shaped 8K capacity run completed 40/40 at concurrency eight. Its 0.766/1.075 s TTFT p50/p95 and 0.906/1.148 s E2E p50/p95 are bounded request-completion measurements. The reported 17.85 aggregate tok/s used only 86 output tokens and is not a controlled decode rate.

Quality¶

The repeated quality run passed context, tools, session recall, and timeout triage. Strict unified-diff formatting passed 2/3 because one answer included an extra leading space after the diff marker. The separate 65.72% GPQA artifact used the pre-final image and is supporting history, not a final-image rerun.

Decision and promotion state¶

The exact configuration was human-approved as Heavy and later declared a pinned rollback in its dated campaign. Retain that historical decision and the complete recipe without presenting it as the current rollback chain or a live route assignment.

Failures and gotchas¶

Engine and EOS coupling¶

Do not substitute a stock engine or silently retag. The checkpoint varies expert counts and attention windows by layer, and its generation config omits the Harmony <|call|> EOS token required by the working tool path. Preserve the exact fork commit, full generation-config override, Marlin backend, and V2-runner disable.

Quality boundary¶

Strict unified-diff formatting passed only 2/3. The earlier 65.72% GPQA artifact used the pre-final image and was not rerun on the final qualified image. No general superiority over Gemma is claimed.