# Qwen3.8-27B
Qwen3.8-27B is the dense, natively multimodal 27B open-weight model of the [[Qwen 3.8]] generation, released on Hugging Face on 5 August 2026 under Apache 2.0 (an FP8 checkpoint followed on 13 August). It is the successor to [[Qwen3.6-27B]] in the same slot: the biggest model that fits on one 24 GB consumer GPU at 4-bit, and the one most people running [[Qwen]] locally will actually use.
Alibaba calls the 3.8 generation "the most capable in the Qwen open-model family to date", built on the Qwen 3.5 architecture. The 27B carries the same gains as [[Qwen 3.8|Qwen3.8-Max]] down to laptop scale: coding, professional work, research, long-horizon agentic tasks, and vision (images, documents, hour-scale video).
## Architecture
- Dense 27B causal LM with a vision encoder (no [[AI Mixture of Experts (MoE)|MoE]]; every parameter is active)
- 64 layers in a hybrid layout: 16 blocks of three Gated DeltaNet (linear attention) layers followed by one Gated Attention layer. Only a quarter of the layers do full attention, which is where the long-context memory savings come from
- Multi-token prediction head, trained with multiple steps, so speculative decoding works out of the box
- Context: 262,144 tokens natively, extensible to 1M
- Thinking on by default, switchable per request; `reasoning_effort` in `low`, `medium`, `xhigh` (the default); `preserve_thinking` keeps reasoning across turns in agentic sessions
- Recommended sampling: thinking mode `temperature=1.0, top_p=0.95, top_k=20`; instruct mode `temperature=0.7, top_p=0.8, presence_penalty=1.5`
A note from the model card worth remembering for any agentic model: **lower reasoning effort does not always reduce total task time**. Faster turns, less analysis, more failed steps, more retries. Total tokens can go up.
## Benchmarks
Alibaba's table, mostly measured through the [[Claude Code]] harness at 256K context:
| Benchmark | Qwen3.6-27B | Qwen3.7-Plus | **Qwen3.8-27B** | Opus 4.6 Max |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 63.4 | 64.0 | **73.0** | 78.2 |
| SWE-bench Pro | 53.5 | 57.6 | **61.7** | 53.4 |
| DeepSWE 1.1 | 13.3 | 14.2 | **42.2** | |
| NL2Repo-Bench | 36.2 | 41.1 | **42.3** | 47.6 |
| CoWorkBench | 61.0 | 65.1 | **70.7** | 68.2 |
| Agents' Last Exam (score) | 27.3 | 33.6 | **42.9** | |
| IFBench | 69.1 | 79.1 | **79.5** | 62.5 |
| GPQA Diamond | 87.8 | 90.3 | **89.2** | 91.3 |
| HLE | 24.0 | 34.7 | **30.8** | 40.0 |
| LiveCodeBench v6 | 83.9 | 89.6 | **90.3** | 88.8 |
| OSWorld-Verified | 63.9 | 73.3 | **84.3** | 72.7 |
| AndroidWorld | 70.3 | 81.0 | **81.9** | 62.0 |
Two things stand out. DeepSWE tripled over 3.6-27B (13.3 → 42.2), which is the agentic-coding jump the whole generation is built around. And a 27B dense model beats Alibaba's own 397B-total Qwen3.7-Plus on most agentic rows. The reasoning rows (GPQA, HLE) are where it still trails; that's the price of 27B.
Usual caveats: Alibaba's numbers, several in-house benchmarks (QwenSWEBench, CoWorkBench, RecreationBench), and the Opus 4.6 column is an older Anthropic model.
## Running it locally
The question that matters for a 27B is which quantization. Quesma benchmarked the Unsloth quants on Terminal-Bench 2.1, GPQA Diamond, and IFBench, which is the first careful measurement of this I've seen for a current model:
| Quant | Size | Terminal-Bench 2.1 | GPQA Diamond |
|---|---|---|---|
| BF16 | 55 GB | ~75% | baseline |
| Q8_0 | 29 GB | ~75% | ≈ baseline |
| Q4_K_M | 17 GB | ~75% | ≈ baseline |
| UD-Q2_K_XL | 10.7 GB | ~70% | modest drop |
| UD-IQ1_S | 6.2 GB | collapses | ~random (25%) |
**4-bit holds up, 2-bit costs a few points, 1-bit is unusable.** IFBench showed no difference at all down to 2-bit. One subtlety from the HN thread: the 2-bit model needed about 25% more output tokens on the tasks it did solve, and a commenter's theory is that a quantized Qwen 3.8 compensates for a worse next-token distribution by thinking longer at `xhigh`. Same success rate, more time. Their recommendation, which I'd repeat: pick the best model that fits in VRAM alongside the context you need; Q4_K_M is the default.
Practical reports from the thread:
- RTX 3090 with Q4_K_M: usable for agentic work, though [[Qwen3.6-35B-A3B]] is much faster at the same quant (3B active vs 27B)
- 16 GB cards (5060 Ti, 9070 XT): Q3 quants with 32K to 87K context at 30 to 50 tokens/s; the 4-bit quants don't leave room for useful context on 16 GB. A dynamic 3-bit quant from ISTA-DASLab (GSQ-RCO) is worth testing there
- Apple Silicon: an M1 Max 64 GB gets ~30 tokens/s with MLX-based servers, much less with `llama-server`. At `xhigh` it thinks long enough that a task takes 3× a cloud model, so people run it overnight or treat it like image generation: send, come back in 40 minutes
- KV cache: q8_0 KV is what fits the model plus 100K context into 24 GB; nobody has measured what that costs in quality yet
- A well-tuned `llama-server` config with MTP draft plus n-gram speculative decoding gets ~70% draft acceptance on code and 30 to 50 tokens/s on a 5060 Ti
And the economic reality check, also from the thread: a 284B [[DeepSeek v4|DeepSeek V4 Flash]] costs about $0.10 per million output tokens from the cheapest hosts, while renting a GPU to run the 27B yourself lands at $3 to $5 per million. Dense 27B needs roughly twice the FLOPs per token of a 13B-active MoE. Local only wins on privacy, on things cloud models refuse, or when the hardware is already paid for.
## References
- Model card: https://huggingface.co/Qwen/Qwen3.8-27B (FP8: https://huggingface.co/Qwen/Qwen3.8-27B-FP8)
- Quesma, "Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses": https://quesma.com/blog/qwen38-27b-quantizations-benchmarked/
- [[Hacker News]] discussion of the quantization benchmark (285 points): https://news.ycombinator.com/item?id=49611128
- Unsloth quants: https://huggingface.co/unsloth/Qwen3.8-27B-GGUF
- Serving recipes: SGLang https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.8-27B · vLLM https://recipes.vllm.ai/Qwen/Qwen3.8-27B
## Related
- [[Qwen]]
- [[Qwen 3.8]]
- [[Qwen3.8-Flash-Next]]
- [[Qwen3.6-27B]]
- [[Qwen3.6-35B-A3B]]
- [[Unsloth]]
- [[llama.cpp]]
- [[vLLM]]
- [[Ollama]]
- [[LM Studio]]
- [[AI Open Weight Models]]
- [[AI Mixture of Experts (MoE)]]
- [[AI Reasoning Models]]
- [[Context Window]]
- [[DeepSeek v4]]
- [[Claude Code]]
- [[Large Language Models (LLMs)]]
- [[Hacker News]]