Agents-A1-NVFP4

Agents-A1 — NVFP4 (W4A4, calibrated MoE)

Calibrated NVFP4 quantization of InternScience/Agents-A1 (agentic Qwen3.5-35B-A3B hybrid MoE) for vLLM. 21.8 GB — and it matches or beats the official FP8 on every quality axis we measured, at 60% of its size.

Quantized and gate-verified by protoLabs on 2× RTX PRO 6000 Blackwell (sm120), vLLM 0.22.1.

Quality — paired gate vs InternScience/Agents-A1-FP8 (official)

Same suites, same judge, same harness, thinking-on, 8192-token budget both sides (pairing across budget methodologies is how quant evals lie — we re-baselined the FP8 ourselves rather than quote its unbounded-budget numbers):

axis                                FP8 official   NVFP4 (this)   delta
----------------------------------  ------------   ------------   ------
reasoning-v2 (24, solver-graded)    0.840          0.868          +0.028
function-call (54, deterministic)   88.9%          90.7%          +1.8
claw agentic (paired-85, judged)    0.640          0.648          +0.008
code spec-delta (8, exec, x3)       0.476 ±0.098   0.453 ±0.072   −0.023

Long-context coherence: adversarially probed at 4K/16K/32K/60K — needle recall perfect, zero degeneration flags, hostile-judge clean. Note for near-max-context use: this model reasons at length — at 60K depth, leave >5K tokens of headroom or disable thinking (chat_template_kwargs: {"enable_thinking": false}), or generation hits the context ceiling mid-think.

Speed — single RTX PRO 6000, client-side seeded benchmark

regime (ISL/OSL)   C    output tok/s   TTFT p50   goodput (TTFT≤2s, TPOT≤50ms)
----------------  ---   ------------   --------   ----------------------------
chat 1k/1k          1        215           63ms    —
chat 1k/1k          8       1028          300ms    1.00
context 8k/1k       1        200          313ms    —
context 8k/1k       8        751         1630ms    0.63

A 35B-class agentic model at 215 tok/s single-stream on one workstation card.

Serving (the config that works on sm120 — not optional)

VLLM_USE_FLASHINFER_SAMPLER=0 vllm serve protoLabsAI/Agents-A1-NVFP4 \
  --moe-backend marlin --max-num-seqs 256 --language-model-only \
  --reasoning-parser qwen3 --tool-call-parser qwen3_coder --enable-auto-tool-choice

• --moe-backend marlin is required on sm120 (RTX 50xx / PRO 6000): the default flashinfer/trtllm fused-MoE FP4 kernel segfaults in its Sm120_SafeFP4 path (upstream report with native trace forthcoming; cross-ref flashinfer#3119).
• --max-num-seqs 256: the hybrid DeltaNet state cache scales with max sequences.
• --language-model-only: text-only serving; the vision-tower profile pass crashes on this stack. (Vision weights ship in the checkpoint, unquantized, for stacks that handle them.)
• MTP tensors are included (grafted from the Qwen3.5-35B-A3B base via save_mtp_tensors_to_checkpoint — Agents-A1 shipped without them), but vLLM spec-decode is currently incompatible with the marlin MoE backend (the global backend must also serve the bf16 draft MoE). For MTP speedups use the GGUF build.

Recipe (provenance — reproduce it)

• llm-compressor main @0.12.1.dev41 (post-#2848 — release 0.12.0 lacks qwen3.5-MoE support entirely; builds before 2026-06-22 save a broken expert layout), transformers==5.12.2 (exactly — 5.10 and ≥5.13 both break this path differently), compressed-tensors 0.17.1.
• NVFP4 W4A4, 128 × ultrachat @2048, moe_calibrate_all_experts=True.
• Kept bf16: MoE router gates + shared-expert gate (low-bit routers corrupt MoE), all linear_attn.* (DeltaNet), vision tower, lm_head, embeddings, MTP tensors.
• Pipeline: protoLabsAI/protoLab → experiments/quantize/a1_requant.py.

Need a different quant?

Open a Community discussion — requests usually ship within 48h. All benchmark rows behind this card (both sides of the pairing, including where we lose): protoLabsAI/lab-benchmarks · charts at protolabs.studio/lab.