Agents-A1¶

Current status and review date¶

Decision snapshot

Product role: historical human-approved Primary profile with retained text, multimodal, router, and promotion-era evidence; it is not presented as the current Primary or as a claim about live routing.
Selected or best-qualified configuration: official FP8 with thinking disabled, FP8 KV, 262,144-token context, concurrency one, and direct text/image/video support.
Measured hardware: one NVIDIA RTX PRO 6000 Blackwell Max-Q on Primary Node.
Evidence: 240K retrieval, tools 20/20, 28/30 strict multimodal cases, and 155.8 tok/s decode at 231,426 actual prompt tokens.
Decision: retain the official FP8 profile as reproducible promotion-era evidence; BF16 remains the correctness control and ProtoLabs NVFP4 remains a text-only no-promotion alternative.
Important limitation: thinking must remain disabled for the qualified contract, and the two retained video failures prevent a perfect strict multimodal result.
Review dates: retained evidence through 2026-07-29; dossier-format review 2026-08-31.

Review narrative¶

2026-07-27–28 — multimodal and quantization qualification¶

BF16 and official FP8 both completed the retained strict multimodal corpus at 28/30. ProtoLabs NVFP4 qualified as a compact text-only option, while its publisher-documented vision-tower crash kept image and video outside that profile. The generated FP8 MoE tune loaded successfully but regressed the three-run 8K c16 throughput mean by 1.399%, so the default kernel selection remained the qualified recipe.

2026-07-29 — 262K head-to-head and promotion-era decision¶

The official FP8 profile passed a 262,144-token operational context at c1, including 240K retrieval and tools 20/20, then became the human-approved Primary through the managed transaction. That is a dated product decision; the profile is now retained as historical evidence rather than described as a current live assignment.

Immutable identity¶

BF16: InternScience/Agents-A1 revision addff08f1653ee72765c5cf458fe84556bb34f8e.
Official FP8: InternScience/Agents-A1-FP8 revision 4d7d59380f327b76e73bc71f40e0c589ad0ca1d5.
ProtoLabs NVFP4: protoLabsAI/Agents-A1-NVFP4 revision ff24ba5c35b99af25d7bf03c7997be5a0d2a5520.
Qualified vLLM revision: f25953cc59f9b4ba9b04b16228d2b86dcfbcbdb1; promotion-era image digest sha256:212a1bd7b4267c604408d17dc0048ef152101bc67fbe6ba8567899fc1f227bcd.

Tested hardware and topology¶

One RTX PRO 6000 Blackwell Max-Q on Primary Node. Qualification used isolated candidate serves; the exact official FP8 profile was promoted through the managed Primary transaction on 2026-07-29. That historical promotion does not describe current live serve or route state.

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

Official FP8 262K profile¶

The qualified profile uses pinned vLLM nightly f25953cc59f9b4ba9b04b16228d2b86dcfbcbdb1, official FP8 weights, FP8 KV, thinking disabled, 262,144 configured tokens, and concurrency one. The exact managed recipe is retained in configs/agents-a1-qwen-262k-head-to-head-recipes.toml.

BF16 and earlier concurrency controls¶

The earlier 131,072-token BF16 multimodal profile reached c16. The official FP8 profile reached c32 at that earlier window. Their text and multimodal recipes remain in configs/agents-a1-qualification-text-recipes.toml and configs/agents-a1-qualification-multimodal-recipes.toml.

ProtoLabs NVFP4 and rejected tune¶

NVFP4 uses text-only Marlin with the FlashInfer sampler disabled and no MTP; its compact recipe is configs/agents-a1-qualification-nvfp4-compact-recipe.toml. The hardware-specific official FP8 MoE tune was generated and loaded but is rejected; the default kernel selection remains the qualified recipe.

Evidence by measurement class¶

Functional and quality¶

functional, capacity, and quality evidence covers BF16 and FP8 text, image, OCR, direct video, tools, streaming, Responses, session, unified-diff, timeout, and 128K c1/c2/c4 gates. Both scored the same 28/30 multimodal corpus result: 12/12 image, 4/4 mixed, and 12/14 video. NVFP4 passed the repeated text gate, 128K c4, and a compact-allocation follow-up.

Capacity and performance¶

At 262K, official FP8 passed 240K retrieval and 20/20 tools, retained 51.93 GiB KV, and measured 32.97 s TTFT plus 155.8 tok/s decode at 231,426 prompt tokens. Client-observed effective prefill includes queueing and scheduling; a kernel-only prefill rate was not measured.

Router isolation¶

The isolated FP8 router passed same-dialect video, media admission, tools, SSE, malformed media, and fail-closed unsupported-dialect probes after the two recorded router fixes.

Decision and promotion state¶

Historical official FP8 decision¶

Official FP8 was the human-gated current Primary in the dated 2026-07-29 campaign after winning the bounded Qwen comparison and passing the complete three-repetition protocol-v3 suite at the 262K serving profile. It is retained as historical promotion-era evidence and is not a current live-state claim.

Retained controls¶

BF16 is the correctness control. NVFP4 is Pareto-preferred for compact text-only deployment, not image/video, and remains no-promotion. The generated FP8 MoE tune is not adopted: its three-run 8K c16 throughput mean regressed 1.399% and missed the 5% gate.

Failures and gotchas¶

Thinking and multimodal assertion boundary¶

Thinking must remain disabled for the qualified contract. The two video failures localized the exact event interval but omitted one required assertion word; identical BF16 and FP8 output rules out a measured FP8 regression. NVFP4's publisher documents a vision-tower crash.

Context and timing interpretation¶

The earlier 131K profiles reject 240K; official FP8 accepts it only in the 262K c1 profile. Client-observed effective prefill includes queueing and scheduling and must not be read as a kernel-only prefill rate.