Agents-A1 FP8 Primary promotion¶

Observed: 2026-07-29

Host: Primary Node, one NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition (96 GB, sm_120)

Decision: promote Agents-A1 official FP8 to Primary with thinking disabled; retain Qwen3.5 122B NVFP4 as the immediate managed rollback.

Why the decision changed¶

The earlier 262K head-to-head established that Agents-A1 used 51.78% less model memory than Qwen, retained 3.75 times as much KV memory, cut measured 240K TTFT by 52.15%, decoded 2.59 times faster at that lane, and was the only exact runtime that delivered the video corpus. That campaign deliberately stopped short of promotion because Agents-A1 had not repeated the complete protocol-v3 quality suite at 262K.

The missing suite was rerun three times with thinking disabled and passed every hard serving-contract assertion. The operator then supplied the separate human promotion authorization. This is an operational Primary decision, not a claim of general intelligence superiority.

Exact promoted profile¶

Field	Value
Model	`InternScience/Agents-A1-FP8`
Revision	`4d7d59380f327b76e73bc71f40e0c589ad0ca1d5`
Served name	`agents-a1-fp8-mm-262k`
Engine	vLLM `0.23.1rc1.dev1327+gf25953cc5`
Image	`vllm/vllm-openai:nightly-f25953cc59f9b4ba9b04b16228d2b86dcfbcbdb1`
Image digest	`sha256:212a1bd7b4267c604408d17dc0048ef152101bc67fbe6ba8567899fc1f227bcd`
Quantization / KV	compressed-tensors FP8 / FP8
Context / admission	262,144 tokens / one sequence
Media limits	four images and one video
Thinking contract	disabled and not caller-overridable at the router

The generated hardware-specific MoE tune remains rejected and inert. Its matched three-run A/B regressed primary-lane throughput by 1.399%, so the promoted recipe keeps vLLM's default selection.

Qualification gate¶

The 240K functional preflight passed smoke, deterministic JSON, retrieval, and 20/20 tool calls. The full protocol-v3 run then passed at a required 100% rate:

Gate	Result
32K, 128K, and 240K context retrieval	3/3 lanes
Automatic tool calling	3/3
Session recall	3/3
Unified diff	3/3
Timeout triage	3/3
Visible answers / allowed finish state	all pass
Reasoning leakage while disabled	none

The repeated run used 256 visible-answer tokens, zero reasoning headroom, three repetitions, and the exact thinking-disabled control artifact. Client-observed timing was:

Context lane	TTFT	End to end
32K	1.571 s	1.918 s
128K	12.839 s	13.212 s
240K	35.209 s	35.615 s

Across the suite, derived decode throughput was 169.19 tok/s p50 and 181.44 tok/s p95; effective prefill was 10,601 tok/s p50. These are client-observed request metrics, not isolated kernel measurements.

Multimodal caveat¶

The retained strict corpus result remains 28/30 for both BF16 and official FP8: 12/12 images, 4/4 mixed-media attempts, and 12/14 video attempts. Both failed event-localization assertions named the exact time interval but omitted the required word alert; matched BF16 output means this was not an observed FP8 quantization regression. The promotion accepts that documented prompt- contract caveat. Direct and isolated routed video_url handling, malformed media, media admission, tools with video, and streaming video were already verified.

Managed transition and rollback¶

The qualification container was removed after an exact identity check. serves promote agents-a1-fp8-primary then quiesced and drained the live Primary tier, started the exact promoted recipe, reran its thinking-disabled functional gate, installed the pinned router configuration, checked exact model identity, and readmitted the tier. Qwen3.5 remains available through the primary-qwen35-rollback managed service and matching rollback router config. The post-promotion llm.primary route then passed smoke, deterministic JSON, and 20/20 tool calls with no reasoning leakage.

No raw Docker lifecycle command was used. Promotion evidence and the terminal serve/router state are retained with the raw benchmark artifacts.