Qwen 3.8 27B (Qwen3.8-27B) is a dense 27B vision-language model under Apache-2.0. Qwen reports 61.7 on SWE-bench Pro and 73.0 on Terminal-Bench 2.1, up from 53.5 and 63.4 for Qwen3.6-27B at the same size. Native context is 262,144 tokens and extends to 1M. The BF16 weights take about 54 GB, so it runs on a single 80 GB GPU.
Qwen 3.8 27B in Short
Last updated October 26, 2026.
- Size: 27B dense with a vision encoder for images and video.
- Context: 262,144 tokens native, extensible to 1,000,000.
- License: Apache-2.0.
- Release: Hugging Face and ModelScope on August 14, 2026, per the QwenLM README.
- Coding: 61.7 on SWE-bench Pro and 73.0 on Terminal-Bench 2.1 in Qwen's runs.
- Serving: Qwen publishes vLLM and SGLang recipes. Ollama lists it as qwen3.8:27b.
Qwen 3.8 is a family. The 27B is the one most people mean by "Qwen 3.8" because it fits on one GPU. The flagship, Qwen3.8-2.4T-A95B, has 2.4T total parameters with 95B active and ships under a custom Qwen license instead of Apache-2.0. For where the 27B sits among every open-weight release, see the best open source LLM ranking.
Qwen 3.8 27B Architecture
Qwen3.8-27B keeps the hybrid layout Qwen introduced with Qwen3.5. Three of every four layers use Gated DeltaNet, a linear-attention block whose cost grows linearly with sequence length. The fourth uses gated full attention. With full attention in only a quarter of the layers, a 262K window needs far less KV cache than a pure-attention 27B would.
| Component | Value |
|---|---|
| Parameters | 27B (dense) |
| Layers | 64: 16 x (3 x Gated DeltaNet + 1 x Gated Attention) |
| Hidden dimension | 5,120 |
| Gated DeltaNet heads | 48 for V, 16 for QK, head dim 128 |
| Gated Attention heads | 24 for Q, 4 for KV, head dim 256 |
| FFN intermediate | 17,408 |
| Vocabulary | 248,320 (padded) |
| Multi-token prediction | Trained with multiple steps |
| Context | 262,144 native, up to 1,000,000 |
Qwen 3.8 27B Benchmarks
All numbers below come from Qwen's model card. Qwen ran SWE-bench Pro, NL2Repo-Bench, DeepSWE 1.1, and QwenSWEBench in the Claude Code harness with a 256K window. The Opus 4.6 Max SWE-bench Pro score is Anthropic's published figure. QwenSWEBench and CoWorkBench are Qwen in-house benchmarks.
| Benchmark | Qwen3.8-27B | Qwen3.6-27B | Muse Glimmer-30B | Opus 4.6 Max |
|---|---|---|---|---|
| SWE-bench Pro | 61.7 | 53.5 | 51.2 | 53.4 |
| Terminal-Bench 2.1 (Terminus) | 73.0 | 63.4 | 51.7 | 78.2 |
| DeepSWE 1.1 | 42.2 | 13.3 | -- | -- |
| NL2Repo-Bench | 42.3 | 36.2 | -- | 47.6 |
| QwenSWEBench (in-house) | 79.0 | 49.3 | -- | 63.8 |
| LiveCodeBench v6 | 90.3 | 83.9 | -- | 88.8 |
| GPQA Diamond | 89.2 | 87.8 | 83.5 | 91.3 |
| HLE | 30.8 | 24.0 | 22.0 | 40.0 |
| IFBench | 79.5 | 69.1 | 77.0 | 62.5 |
| CoWorkBench (in-house) | 70.7 | 61.0 | -- | 68.2 |
Opus 4.6 Max still leads on Terminal-Bench 2.1, NL2Repo-Bench, GPQA Diamond, and HLE. On SWE-bench Pro the 27B beats it in Qwen's table, but Qwen re-ran its own baselines on a corrected task set while Opus keeps its official score, so the two numbers were not measured the same way.
Vision and Computer-Use Benchmarks
Qwen3.8-27B is natively multimodal. The largest gains over Qwen3.6-27B are on agent tasks that act through a screen.
| Benchmark | Qwen3.8-27B | Qwen3.6-27B | Opus 4.6 Max |
|---|---|---|---|
| OSWorld-Verified (computer use) | 84.3 | 63.9 | 72.7 |
| WebArena-Verified (browser use) | 64.8 | 48.8 | -- |
| AndroidWorld (mobile use) | 81.9 | 70.3 | 62.0 |
| SWE-MM (multimodal SWE) | 38.6 | 25.7 | 27.1 |
| Vision2Web | 62.9 | 45.0 | -- |
| OmniDocBench 1.5 | 91.1 | 89.4 | 86.6 |
Qwen 3.8 27B VRAM Requirements
Weight memory is parameters times bytes per parameter. For 27B parameters that is about 54 GB at BF16, 27 GB at 8-bit, and roughly 14 GB at 4-bit. The KV cache is extra and grows with context length and batch size. Only one layer in four keeps a full-attention KV cache, so long contexts cost less memory than on a dense transformer of the same size.
| Precision | Weights | Fits on |
|---|---|---|
| BF16 | ~54 GB | One 80 GB GPU (H100, A100 80GB) |
| 8-bit (FP8 / Q8) | ~27 GB | One 48 GB GPU, or a 32 GB card with short context |
| 4-bit (Q4 GGUF) | ~14 GB | 24 GB consumer GPUs, Apple silicon with 32 GB+ |
Pick the precision whose weights leave at least a quarter of the card free. That headroom holds the KV cache for a useful agent context.
For the GPU side of this math, see NVIDIA H200 pricing and specs.
How to Run Qwen 3.8 27B
Qwen lists SGLang, vLLM, and TokenSpeed recipes on the model card and recommends the latest version of each. The examples below start an OpenAI-compatible server with the standard CLI of each engine.
vLLM
vllm serve Qwen/Qwen3.8-27B --max-model-len 262144SGLang
python -m sglang.launch_server --model-path Qwen/Qwen3.8-27B --context-length 262144Ollama
ollama run qwen3.8:27bSampling matters more than usual. In thinking mode Qwen recommends temperature 1.0, top_p 0.95, and top_k 20. In non-thinking mode it recommends temperature 0.7, top_p 0.8, top_k 20, and presence_penalty 1.5. The reasoning_effort parameter trades reasoning depth for cost, and preserve_thinking is on by default.
Qwen 3.8 27B vs Qwen 3.6 27B
Both models are dense 27B checkpoints under Apache-2.0 on the same hybrid attention base, and the upgrade is in post-training. In Qwen's table the 27B gains 8.2 points on SWE-bench Pro, 9.6 on Terminal-Bench 2.1, 28.9 on DeepSWE 1.1, and 20.4 on OSWorld-Verified. If you already serve Qwen 3.6, the swap needs no new hardware. Re-tune sampling, because thinking is on by default.
Qwen 3.8 License
The Qwen3.8-27B repository ships the standard Apache License 2.0. That grants copyright and patent licenses for commercial use, modification, and redistribution. The 2.4T flagship, Qwen3.8-2.4T-A95B, uses its own Qwen license, so read that file before you host it.
Pros and Cons
- 61.7 SWE-bench Pro from a dense 27B in Qwen's harness
- Apache-2.0 weights
- BF16 weights fit one 80 GB GPU
- 262K native context, extensible to 1M
- Image and video input in the same checkpoint
- Coding scores are vendor-run on a corrected task set
- Thinking on by default raises output tokens
- Trails Opus 4.6 Max on Terminal-Bench 2.1 and HLE
- The 2.4T flagship is not Apache-2.0
Frequently Asked Questions
What is Qwen 3.8 27B?
A dense 27B vision-language model from Alibaba's Qwen team, on Hugging Face since August 14, 2026 under Apache-2.0, with 262,144 tokens of native context.
How good is Qwen 3.8 27B at coding?
Qwen reports 61.7 on SWE-bench Pro and 73.0 on Terminal-Bench 2.1, against 53.5 and 63.4 for Qwen3.6-27B.
How much VRAM does Qwen 3.8 27B need?
About 54 GB of weights at BF16, 27 GB at 8-bit, and 14 GB at 4-bit, plus the KV cache.
Is Qwen 3.8 open source?
The 27B is Apache-2.0 open weights. The training data is not released.
How do I turn off thinking in Qwen 3.8?
Disable it per request. The reasoning_effort parameter tunes depth when thinking is on.
Should I upgrade from Qwen 3.6 27B?
For agent work, yes. The hardware is the same and the agentic scores rose by 8 to 29 points in Qwen's table.
The fastest endpoints are private deployments
Morph's top speeds come from dedicated deployments, not shared public endpoints: speculators trained on your traffic, caching tuned to your workload, and volume discounts over public per-token rates. Over 500 billion tokens per day run this way.
Give a Qwen 3.8 agent the right code
WarpGrep searches a repository in parallel and returns only the relevant files, so a 27B model spends its context on the code that matters.