Qwen3.8-27B
Qwen3.8-27B is the mid-size dense model of Alibaba's August 2026 Qwen3.8 release, under Apache 2.0. It mixes Gated DeltaNet linear-attention layers with gated full-attention layers and includes a vision encoder for image and video input. Thinking is on by default but can be turned off per request.
Specifications
- Context window
- 256Kabout 197,000 words, or 390 pages
- Max output
- Not published
- Input price
- $0.80per 1M tokens
- Output price
- $4per 1M tokens
- Cached input
- Not published
- Weights
- OpenApache-2.0
- Parameters
- 27.8Bdense
- Input types
- Text, Images, Video
- Output types
- Text
- Reasoning mode
- Yes
- Tool calling
- Yes
- Released
- Aug 14, 2026Current
Groq price (qwen/qwen3.8-27b, a preview model on Groq, served with a 131,072-token context and 16,384 max completion tokens). Official context is 262,144 native, extensible to 1M with YaRN.
What it costs in practice
| Workload | Monthly cost at list price |
|---|---|
| Support chatbot30,000 conversations a month, 2,000 input and 400 output tokens each | $96 |
| Document summaries2,000 documents a month, 12,000 input and 800 output tokens each | $26 |
| Coding agent1,000 tasks a month, 60,000 input and 6,000 output tokens each | $72 |
Standard rates, no caching or batch discounts. Run your own numbers.
Strengths
- Apache 2.0 with image and video input
- Only 16 of 64 layers keep a full KV cache
- About 14 GB of weights at 4-bit, within reach of a 24 GB GPU
Trade-offs to weigh
- Thinking mode is on by default and adds output tokens
- Few hosts list it yet; Groq offers it only as a preview
Where it fits
Alibaba (Qwen) positions Qwen3.8-27B for:
Where you can use it
- Hugging Face
- Ollama
- Groq
Run it yourself
Released under Apache-2.0. Approximate memory needed just to load the weights:
| Precision | Memory for weights |
|---|---|
| 16-bit (BF16/FP16) | 56 GB |
| 8-bit (FP8/INT8) | 28 GB |
| 4-bit (Q4_K_M, about 4.85 bits) | 17 GB |
Add memory for the KV cache (it grows with context length and concurrent users) and 1-2 GB of runtime overhead. Size it precisely in the VRAM calculator.
Alternatives to consider
| Model | Tier | Context | Input / output per 1M | |
|---|---|---|---|---|
| Grok 4.3xAI | Efficient | 1M | $1.25 / $2.50 | Compare |
| Grok Build 0.1xAI | Efficient | 250K | $1 / $2 | Compare |
| Muse Glimmer 30BMeta | Efficient | 128K | $0.35 / $1.50 | Compare |
| Gemini 3.5 Flash-LiteGoogle | Efficient | 1M | $0.30 / $2.50 | Compare |
Frequently asked questions
How much does Qwen3.8-27B cost?
Qwen3.8-27B lists at $0.80 per million input tokens and $4 per million output tokens (checked October 10, 2026). Groq price (qwen/qwen3.8-27b, a preview model on Groq, served with a 131,072-token context and 16,384 max completion tokens). Official context is 262,144 native, extensible to 1M with YaRN. Confirm current rates on the provider's pricing page before you commit.
What is the context window of Qwen3.8-27B?
Qwen3.8-27B accepts up to 262,144 tokens of context (about 197,000 words, or 390 pages). Long prompts cost more and can be slower, so most apps send only what the task needs.
Is Qwen3.8-27B open source?
Qwen3.8-27B is an open-weight model released under the Apache-2.0 license. You can download the weights from Hugging Face (Qwen/Qwen3.8-27B) and run or fine-tune them yourself, within the license terms. Open weights isn't the same as fully open source: training data and code may not be published.
Can I run Qwen3.8-27B locally?
Yes, if you have enough memory. The weights alone need about 17 GB at 4-bit quantization and about 56 GB at 16-bit, plus room for the KV cache. Use the VRAM calculator on LLModel.com to check a specific GPU or Mac.
What is Qwen3.8-27B best for?
Alibaba (Qwen) positions Qwen3.8-27B for self-hosting, coding, agents, multimodal. It accepts text, images, video as input, has a reasoning (thinking) mode and supports tool calling. Test it on a sample of your own tasks before you choose it for production.
Sources
- huggingface.co/Qwen/Qwen3.8-27B
- console.groq.com/docs/models
- github.com/QwenLM/Qwen3.8
- huggingface.co/Qwen/Qwen3.8-27B/raw/main/config.json
- huggingface.co/Qwen/Qwen3.8-27B/raw/main/chat_template.jinja
- huggingface.co/api/models/Qwen/Qwen3.8-27B
- ollama.com/library
Facts checked against these pages on October 10, 2026. Spot an error? Tell us.
Still choosing a model?
Send us your use case and constraints. We'll come back with a shortlist and an evaluation plan you can run on your own data.
- Free and vendor-neutral: no model provider pays for placement.
- Your details go only to partners you agree to.