Ornith-1.0-35B — NVFP4 (W4A4, calibrated MoE)

Calibrated NVFP4 quantization of deepreinforce-ai/Ornith-1.0-35B (Qwen3.5-35B-A3B hybrid MoE) for vLLM. 24 GB, and it decodes at the same speed as FP8 while every quality axis lands at parity or better. This is the model we run as our own daily driver — the gate below is the quant measured against the exact checkpoint serving our production traffic.

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

Quality — paired gate vs protoLabsAI/Ornith-1.0-35B-FP8 (our parity-verified daily driver)

Same suites, same judge, same harness, thinking-on, 8192-token budget both sides. Outliers re-trialed ×3 on both sides before verdicts — one apparent regression (0.88 → 0.00 on a single task) turned out to be baseline drift: the FP8 scores 0.00 on it today too. Quant evals need paired, same-day baselines or they lie.

axis                                FP8 (prod)     NVFP4 (this)   delta
----------------------------------  ------------   ------------   ------
claw agentic (paired-35, judged)    0.672          0.677          +0.005
function-call (54, deterministic)   93.0%          94.4%          +1.4
reasoning-v2 (24, solver, x3)       0.882 ±0.021   0.889 ±0.042   +0.007
code spec-delta (8, exec, x3)       0.616          0.580 ±0.062   −0.036

Long-context coherence: adversarially probed at 4K/16K/32K/60K — needle recall perfect, zero degeneration flags (char-runs, n-gram loops, compression collapse), hostile-judge clean at every rung. Long-context agentic runs verified out to >131K served context.

This is measurement-resolution parity, not a lossless format: deltas straddle zero and sit inside our suites' noise (code-exec single-trial std alone is ±0.06). On dense models NVFP4 costs real quality (see our 9B card: −0.028 claw, one reproducible failure mode); on A3B MoE we cannot distinguish it from the FP8 base.

Speed (speed-test-v2, client-side, random dataset, fixed seed)

regime          C   ttft p50   tpot p50   agg tok/s   goodput
------------   --   --------   --------   ---------   -------
chat 1k/1k      1       62ms      4.8ms       207.8      0.20
chat 1k/1k      8      297ms      7.8ms       950.6      0.81
context 8k/1k   1      305ms      4.9ms       194.2      0.19
context 8k/1k   8     1520ms      9.3ms       738.4      0.64

FP8 reference on the same card: ~207 tok/s at C1 — NVFP4 costs nothing in speed and saves 11 GB of VRAM (more KV headroom, or denser replica packing).

Serving (vLLM ≥ 0.22.1, sm120)

VLLM_USE_FLASHINFER_SAMPLER=0 vllm serve protoLabsAI/Ornith-1.0-35B-NVFP4 \
  --max-model-len 262144 \
  --gpu-memory-utilization 0.90 \
  --max-num-seqs 256 \
  --language-model-only \
  --moe-backend marlin \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_xml --enable-auto-tool-choice

• --moe-backend marlin is required on sm120 — the trtllm auto-selected NVFP4 MoE path (Sm120_SafeFP4) segfaults even on clean checkpoints.
• The checkpoint preserves the MTP speculative-decoding tensors (bf16, excluded from quantization). vLLM's marlin MoE backend can't currently run an unquantized MTP draft over quantized experts, so serve MTP-less; the heads are there for the GGUF pipeline (llama.cpp --spec-type draft-mtp) and for when the backend gap closes.
• DeltaNet linear-attention layers, embeddings, router gates, and lm_head kept bf16 (quantizing any of them breaks the hybrid arch — recipe in repo).

Recipe

llm-compressor (main, post-#2848), NVFP4 (16-elem blocks, E4M3 scales), 128 × 2048-token ultrachat calibration with moe_calibrate_all_experts=True, data-collator pinned for the 4-D DeltaNet inputs, MTP tensors re-attached from the base checkpoint post-save. transformers==5.12.2. Full script in this repo.

---

Need a different size/format? Open a Community discussion — we usually ship within 48h. Numbers on this card trace to rows in protoLabsAI/lab-benchmarks; the board lives at protolabs.studio/lab.