Qwen3.8-27B: The 27B Open-Source Model That Packs Agentic Coding Into Your GPU
Quick verdict
Qwen3.8-27B is Alibaba's ~27B dense open-source model (Apache 2.0) with native image/video understanding and big agentic coding gains — DeepSWE 1.1 at 42.2 and Terminal-Bench 2.1 at 73.0 — available as a ~17GB GGUF that fits on 24GB consumer GPUs.
Alibaba open-sourced the Qwen3.8-27B on August 14, 2026: a ~27 billion parameter dense model that combines native multimodal understanding with surprisingly strong agentic coding — in a size you can actually run on consumer hardware.
For developers who have been waiting for an open model that's both small enough to self-host and smart enough for real coding agents, this is one of the most interesting releases of 2026.
What Is Qwen3.8-27B?
Qwen3.8-27B is part of Alibaba's Qwen3.8 family. It's a dense model (not MoE), designed as the practical, local-friendly counterpart to the huge Qwen3.8-2.4T-A95B open flagship.
Key characteristics:
- ~27B parameters (reported ~27.78B; ~28B in BF16 weights)
- Native multimodal: understands both images and video
- Apache 2.0 open weights
- Flexible thinking control: default thinking on, with
xhigh,medium, andlowreasoning effort levels - 262,144-token native context (262K), extendable to 1M via YaRN
- Hosted version with 1M context and built-in tools is expected soon
Performance: A Big Agentic Leap
The headline is the jump in agentic benchmarks versus the previous Qwen3.6-27B:
| Benchmark | Qwen3.6-27B | Qwen3.8-27B |
|---|---|---|
| DeepSWE 1.1 | 13.3 | 42.2 |
| Terminal-Bench 2.1 | — | 73.0 |
| SWE-bench Pro | — | 61.7 |
According to Qwen's published numbers, Qwen3.8-27B also clears Claude Opus 4.6 Max on SWE-bench Pro, QwenSWEBench, LiveCodeBench v6, OSWorld, and AndroidWorld — despite being a fraction of the size.
⚠️ These are vendor-reported benchmarks. Independent verification — especially of OSWorld/AndroidWorld agent claims — was still pending at the time of writing.
Pricing & Access
- OpenRouter API: ~$0.45 per 1M input tokens, ~$3.20 per 1M output tokens
- Open weights: free to download (Apache 2.0)
- Hosted Qwen version: coming soon, with 1M context and official built-in tools
For API users, it's more expensive than DeepSeek V4 Flash but far cheaper than GPT-5.6-class output.
Local Deployment
This is where the 27B stands out: it's sized for consumer hardware.
- GGUF Q4_K_M quantized model is roughly 17GB
- Reported to run on an RTX 3090 (24GB), dual RTX 5080s, and even 16GB GPUs with tighter quantization
- Native 262K context is a lot for a local model; you'll generally trim context or use YaRN extension depending on your VRAM
Quick practical tips from the local community:
- Choose the right GGUF quantization for your VRAM — Q4_K_M is the popular balance
- Watch the chat template: older UIs often apply the wrong Qwen template, which degrades output
- Use the native 262K window for big-in-context work, and reserve 1M/YaRN for when you really need it
How It Fits the Open-Source Lineup
| Model | Size | Multimodal | Agentic coding strength | Local-friendly |
|---|---|---|---|---|
| Qwen3.8-27B | ~27B dense | Native (image/video) | Very strong for its size | ✅ ~17GB GGUF |
| Qwen3.8-2.4T-A95B | 2.4T MoE (95B active) | Yes | Flagship | ❌ Requires big rigs |
| GLM-5.3 | ~743B MoE | No (text-only) | Very strong | ⚠️ Weights coming |
| DeepSeek V4 Flash | 284B MoE (13B active) | No (text-only) | Strong | ✅ |
| Kimi K3 | 2.8T MoE | Yes | Frontier-level | ❌ Large |
The Qwen3.8-27B is the sweet spot if you want multimodal + agentic coding + local self-hosting in one package.
Who Should Use It
- Developers who self-host: best open multimodal coding model at a size you can run on one consumer GPU
- Agent builders: strong Terminal-Bench/DeepSWE numbers at a scale that's practical to iterate on
- Privacy-conscious teams: Apache 2.0 and local hosting
- Multimodal + coding workflows: native image/video understanding without needing the 2.4T flagship
The Bottom Line
Qwen3.8-27B proves that you don't need a jumbo flagship to get serious agentic coding. For ~17GB of quantized weights, you get native vision, strong coding-agent benchmarks, and a model size that fits a local setup.
If you've been looking for the best open-source multimodal coding model you can run yourself, Qwen3.8-27B is the current leader for the 27B class — just treat the benchmark claims as promising until independent tests land.
Related Articles
Keep reading