nvidia NVFP4 PTQ (BF16→NVFP4)
source ↗MMLU-Pro
85−0.6baseline 85.6
GPQA Diamond
84.8−0.1baseline 84.9
AIME 2025
88.8−0.4baseline 89.2
IFBench
62.8+0.5baseline 62.3
Qwen3.6-35B-A3B
- Organization
- Alibaba
- Parameters
- 35B total / 3B active
- Context
- 262K→1M
- License
- apache-2.0
- Release
- 2026-04
MoE 40 layers, 256 experts top-8+1 shared, Gated DeltaNet
Qwen3.6-35B-A3B is Alibaba's 35B-total / 3B-active MoE with Gated DeltaNet attention, a 262K-to-1M context and an Apache-2.0 license. Its nvidia PTQ NVFP4 build loses under a point on MMLU-Pro, GPQA and AIME, while the NVIDIA QAD W4A4 recipe serves 20,076 tok/s decode — 1.14x BF16 — on 4x GB200.
Qwen3.6-35B-A3B is a 35B total / 3B active apache-2.0 model from Alibaba with 262K→1M context. Quantized checkpoints lose 0.1–4.6 points on the 13 measured per-quant benchmarks.
Base-model benchmarks
SWE-bench Verified
73.4SWE-bench Pro
49.5Terminal-Bench 2.0
51.5MMLU-Pro
85.2MMLU-Redux
93.3GPQA
86HLE
21.4LiveCodeBench v6
80.4AIME26
92.7HMMT Feb25
90.7
Quant checkpoints
| Checkpoint | Family | Size | Source |
|---|---|---|---|
| Qwen3.6-35B-A3B FP8 Official FP8 (35B-A3B); size not captured | FP8 | — | source ↗ |
| nvidia Qwen3.6-35B-A3B-NVFP4 (PTQ) ~3.06x size reduction (PTQ) | NVFP4 | — | source ↗ |
| RedHatAI Qwen3.6-35B-A3B-NVFP4 | NVFP4 | — | source ↗ |
| unsloth Qwen3.6-35B-A3B-GGUF IQ1_M | GGUF | 10 GB | source ↗ |
| unsloth Qwen3.6-35B-A3B-GGUF Q2_K_XL | GGUF | 12.3 GB | source ↗ |
| unsloth Qwen3.6-35B-A3B-GGUF Q3_K_M | GGUF | 16.6 GB | source ↗ |
| unsloth Qwen3.6-35B-A3B-GGUF Q4_K_M | GGUF | 22.1 GB | source ↗ |
| unsloth Qwen3.6-35B-A3B-GGUF Q4_K_XL | GGUF | 22.4 GB | source ↗ |
| unsloth Qwen3.6-35B-A3B-GGUF Q5_K_M | GGUF | 26.5 GB | source ↗ |
| unsloth Qwen3.6-35B-A3B-GGUF Q6_K | GGUF | 29.3 GB | source ↗ |
| unsloth Qwen3.6-35B-A3B-GGUF Q8_0 | GGUF | 36.9 GB | source ↗ |
| unsloth Qwen3.6-35B-A3B BF16 | BF16 | 69.4 GB | source ↗ |
| ISTA-DASLab Qwen3.6-35B-A3B-2Bit-GSQ 2-bit GSQ | GGUF | 12.6 GB | source ↗ |
| palmfuture Qwen3.6-35B-A3B-GPTQ-Int4 | GPTQ | 24.4 GB | source ↗ |
Per-quant quality deltas
RedHatAI NVFP4
source ↗GSM8K-Platinum
96.08+0.35baseline 95.73
GPQA
84.68+0.17baseline 84.51
LiveCodeBench v6
74.67−2.66baseline 77.33
SWE-bench Verified
50.2−4.6baseline 54.8
NVIDIA QAD W4A4 (PTQ)
source ↗MMMU-Pro
73.4−1.2baseline 74.6
IFBench
57.9−2.1baseline 60
NVIDIA QAD W4A4 (+500 iters)
source ↗MMMU-Pro
73.9−0.7baseline 74.6
IFBench
59.6−0.4baseline 60
2-bit GSQ (ISTA-DASLab)
source ↗AIME 2025
93.33+0baseline 93.33
GPQA Diamond
80.3−3.54baseline 83.84
IFEval
87.77−3.48baseline 91.25
MMLU-Pro
81−3.87baseline 84.87
GSM8K
93.93−2.28baseline 96.21
GPTQ Int4 (palmfuture)
source ↗wikitext-2 PPL retention
97.9Serving metrics
What’s not published
- Per-quant GGUF quality is only in image-only KLD plots.
- The three NVFP4 tables use different baseline harnesses, so their rows should not be mixed.
- The 2-bit GSQ and GPTQ Int4 cards use different harnesses from the NVFP4 tables, so their rows should not be mixed.
Sources
- https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8 ↗
- https://huggingface.co/nvidia/Qwen3.6-35B-A3B-NVFP4 ↗
- https://huggingface.co/RedHatAI/Qwen3.6-35B-A3B-NVFP4 ↗
- https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF ↗
- https://nvidia.github.io/Model-Optimizer/announcements/qwen36-w4a4-qad.html ↗
- https://huggingface.co/ISTA-DASLab/Qwen3.6-35B-A3B-2Bit-GSQ ↗
- https://huggingface.co/palmfuture/Qwen3.6-35B-A3B-GPTQ-Int4 ↗
FAQ
- What is Qwen3.6-35B-A3B's license, context window and parameter count?
- apache-2.0 license, 262K→1M context window, 35B total / 3B active parameters (Alibaba).
- How much quality does quantization cost Qwen3.6-35B-A3B?
- Across 17 published deltas, quantized checkpoints differ from the full-precision baseline by 1.47 points on average (range -4.6 to 0.5).
- How many quant checkpoints are cataloged for Qwen3.6-35B-A3B?
- 14 checkpoints across FP8, NVFP4, GGUF, BF16, GPTQ, sizes from 10 GB to 69.4 GB.
- When was Qwen3.6-35B-A3B released?
- 2026-04.
- Where do the Qwen3.6-35B-A3B benchmark numbers come from?
- From 7 official sources — vendor model cards, quantization publisher cards and serving recipe pages.