NVFP4 (nvidia)
source ↗GPQA Diamond
88.01−0.91baseline 88.92
Terminal-Bench 2.1
74.02−1.54baseline 75.56
AA-LCR
73.38+0.75baseline 72.63
MMMU-Pro
74.86−0.28baseline 75.14
SciCode
48.41+0.48baseline 47.93
IFBench
78.93−1.14baseline 80.07
Qwen 3.8 27B
- Organization
- Alibaba
- Parameters
- 27B dense + vision
- Context
- 262K→1M
- License
- apache-2.0
- Release
- 2026-08
dense, 64 layers, Gated DeltaNet + Gated Attention, MTP
Qwen3.8-27B is Alibaba's dense 27B-plus-vision model with 64 layers of Gated DeltaNet and Gated Attention, an MTP module and a 262K-to-1M context, released in August 2026 under the Apache-2.0 license. Its quantization story spans an nvidia NVFP4 checkpoint (18B params) that stays within a point on a six-benchmark table, an ISTA GSQ-RCO 3.5-bit GGUF that is near-lossless, and an AMD Quark MXFP4-AWQ build.
Qwen 3.8 27B is a 27B dense + vision apache-2.0 model from Alibaba with 262K→1M context. Quantized checkpoints lose 0.28–1.54 points on the 5 measured per-quant benchmarks.
Base-model benchmarks
Terminal-Bench 2.1
73SWE-bench Pro
61.7NL2Repo
42.3DeepSWE 1.1
42.2QwenSWEBench
79IFBench
79.5GPQA Diamond
89.2HLE
30.8LiveCodeBench v6
90.3OSWorld-Verified
84.3
Quant checkpoints
| Checkpoint | Family | Size | Source |
|---|---|---|---|
| Qwen Qwen3.8-27B-FP8 ~28 GB official FP8 | FP8 | 28 GB | source ↗ |
| nvidia Qwen3.8-27B-NVFP4 18B params (NVFP4); size not captured | NVFP4 | — | source ↗ |
| ISTA-DASLab Qwen3.8-27B-GSQ-RCO IQ2_S | GGUF | 9.3 GB | source ↗ |
| ISTA-DASLab Qwen3.8-27B-GSQ-RCO IQ3_XXS | GGUF | 10.1 GB | source ↗ |
| ISTA-DASLab Qwen3.8-27B-GSQ-RCO IQ3_S 3.5-bit GSQ-RCO | GGUF | 11.8 GB | source ↗ |
| amd Qwen3.8-27B-Quark-AWQ-MXFP4 16B params (MXFP4-AWQ); size not captured | Quark | — | source ↗ |
| unsloth Qwen3.8-27B-GGUF Q4_K_M | GGUF | 16.5 GB | source ↗ |
| unsloth Qwen3.8-27B-GGUF Q5_K_M | GGUF | 19.8 GB | source ↗ |
| unsloth Qwen3.8-27B-GGUF Q6_K_M | GGUF | 23.1 GB | source ↗ |
| unsloth Qwen3.8-27B-GGUF Q8_0 | GGUF | 29 GB | source ↗ |
| unsloth Qwen3.8-27B-GGUF BF16 | BF16 | 54.7 GB | source ↗ |
Per-quant quality deltas
Serving metrics
What’s not published
- No per-quant quality table is published for the unsloth GGUF ladder (Q4_K_M to Q8_0, BF16).
- No MMLU or AIME numbers appear on the official Qwen FP8 card.
Sources
- https://huggingface.co/Qwen/Qwen3.8-27B-FP8 ↗
- https://huggingface.co/nvidia/Qwen3.8-27B-NVFP4 ↗
- https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF ↗
- https://huggingface.co/amd/Qwen3.8-27B-Quark-AWQ-MXFP4 ↗
- https://huggingface.co/unsloth/Qwen3.8-27B-GGUF ↗
- https://huggingface.co/cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF ↗
FAQ
- What is Qwen3.8-27B's license, context window and parameter count?
- apache-2.0 license, 262K→1M context window, 27B dense + vision parameters (Alibaba).
- How much quality does quantization cost Qwen3.8-27B?
- Across 10 published deltas, quantized checkpoints differ from the full-precision baseline by 0.15 points on average (range -1.54 to 1.67).
- How many quant checkpoints are cataloged for Qwen3.8-27B?
- 11 checkpoints across FP8, NVFP4, GGUF, Quark, BF16, sizes from 9.3 GB to 54.7 GB.
- When was Qwen3.8-27B released?
- 2026-08.
- Where do the Qwen3.8-27B benchmark numbers come from?
- From 6 official sources — vendor model cards, quantization publisher cards and serving recipe pages.