Qwen3.8-27B-NVFP4-all

Qwen3.8-27B-NVFP4-all

+13 to +17% decode throughput on an RTX 5090 over RadixArk/Qwen3.8-27B-NVFP4 at the same GSM8K and MATH-500 accuracy. That is this checkpoint versus the RadixArk export, both served with FlashInfer b12x and the bf16 DFlash2 draft. The full shipping stack against stock SGLang is +25 to +33% on fixed prompts; see thomgardiner/sglang-sm120-nvfp4.

Model Overview

• Model Architecture: Qwen/Qwen3.8-27B, hybrid Gated DeltaNet + full attention, 64 layers
• Input / Output: text / text
• Base checkpoint: RadixArk/Qwen3.8-27B-NVFP4
• Release Date: 2026-09-06
• License: Apache-2.0

The 208 linear projections that RadixArk/Qwen3.8-27B-NVFP4 stores in FP8 are NVFP4 here: linear_attn.in_proj_qkv, in_proj_z, out_proj in the 48 Gated DeltaNet layers and self_attn.q_proj, k_proj, v_proj, o_proj in the 16 attention layers. MLP and lm_head tensors are byte-identical to the base checkpoint. 20.6 GB → 18 GB.

Model Optimizations

Weights are quantized with a symmetric per-block scheme, block size 16, E4M3 block scales and one FP32 global scale per tensor (amax / (448 × 6)). Activations use a static per-tensor scale. Activation scale = the base checkpoint's FP8 static activation scale ÷ 6. No new calibration; the base export was calibrated by Model Optimizer on 1,024 samples of cnn_dailymail at 512 tokens. Tensors that SGLang fuses at load time (in_proj_qkv + in_proj_z, q_proj + k_proj + v_proj) share one global scale per group.

Weights were quantized from the bf16 release of Qwen3.8-27B, not from the FP8 tensors. in_proj_a, in_proj_b, convolutions, norms, embeddings, and the mtp head remain bf16 as in the base checkpoint.

Deployment

python3 -m sglang.launch_server \
  --model-path thomasgardiner/Qwen3.8-27B-NVFP4-all \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path incoai/Qwen3.8-27B-DFlash2 \
  --speculative-num-draft-tokens 8 \
  --attention-backend flashinfer \
  --trust-remote-code \
  --max-running-requests 1 \
  --cuda-graph-max-bs-decode 1 \
  --mem-fraction-static 0.91

Tested with lmsysorg/sglang@sha256:616a3e97f45191af975896cfa644279096cb31bd408a071c2e99ca7209c3cafe (FlashInfer 0.6.17, CUDA 13.0). On SM120, apply the b12x backend patch from thomgardiner/sglang-sm120-nvfp4 so NVFP4 GEMMs run on FlashInfer's SM120 kernel. Without it the GEMMs stay on CUTLASS; see the cutlass row in that repo's serving table.

Creation

python3 convert.py <Qwen/Qwen3.8-27B bf16 dir> <RadixArk/Qwen3.8-27B-NVFP4 dir> <output dir>

convert.py in this repository. Its check mode re-quantizes two of the base checkpoint's NVFP4 MLP tensors from bf16 and compares them to the stored tensors: block scales match 100%, packed weights 99.4 to 99.6%, reconstruction error identical to five digits.

Evaluation

One RTX 5090 per target, SGLang with the b12x patch, DFlash2 bf16 draft, greedy, thinking disabled, 8 concurrent requests. lm-evaluation-harness 0.4, --apply_chat_template --num_fewshot 0, max_gen_toks 4096 (2048 for IFEval and HumanEval).

Every difference is inside one standard error. HumanEval is graded by running the fenced code the model returned against the task's check(); the harness's own humaneval_instruct filter expects a pre-opened code fence and scores a chat model at 0. With thinking enabled, GSM8K on the first 500 items was 83.0 ± 1.7 (RadixArk) and 85.2 ± 1.6 (this checkpoint). Result files and the grader: receipts/lmeval, bench/he_grade.py.

Throughput

Both columns use FlashInfer b12x and the bf16 DFlash2 draft. The change is this checkpoint only.

Named datasets, greedy, thinking on, concurrency 1:

Acceptance length: 3.80 → 3.72 (MT-Bench), 5.07 → 5.06 (GSM8K), 5.14 → 5.05 (MATH-500).

Fixed prompts, greedy, thinking off, streamed verify steps, n=2:

Decode step 18.8 → 16.3 ms on all three. Interleaved n=6 of the shipping stack (this checkpoint + b12x + bf16 draft versus stock CUTLASS + RadixArk) is 156.6 / 346.4 / 353.2 tok/s, +25 / +33 / +26%.

Code/math recipe on this checkpoint: FP8 DFlash2 draft and 12 draft tokens, n=2: 157 / 396 / 439 tok/s. Prose stays at the shipping number. See thomgardiner/sglang-sm120-nvfp4.

One bench_serving run on random-token prompts (ISL 8192 / OSL 1024, concurrency 1, seed 7) read 196.2 → 263.9 tok/s (+34.5%). Acceptance on that shape moves between runs. Do not quote it as the typical gain.

Not measured: batch sizes above 1, HumanEval on this checkpoint, other engines.

Artifact identity

Base: RadixArk/Qwen3.8-27B-NVFP4, conversion-manifest.json carried over unchanged.