What is Qwen3.8-27B?
Qwen3.8-27B is the dense 27B member of Alibaba's Qwen3.8 open family and the successor to Qwen3.6-27B, with the same size and layer layout. It reads text, images and video, thinks before it answers at an effort level you can lower, and keeps its earlier reasoning across turns. In Alibaba's tests it beats the company's own Qwen3.7-Plus on coding and computer-use benchmarks.
You can download the weights free or call it on Qwen Cloud, which serves 1M tokens per request. Qwen Chat and Qwen's Token Plan do not offer it.
What changed from Qwen3.6-27B?
Same size and design, but higher scores, longer thinking by default, a cheaper hosted API and a 1M-token hosted context.
- Same size and layout
- Both are 27B dense models with 64 layers in the same layout, so hardware that runs 3.6 runs 3.8.
- Coding scores
- Alibaba's table moves SWE-bench Pro from 53.5 to 61.7 and Terminal Bench 2.1 from 63.4 to 73.0.
- Reasoning controls
- New xhigh, medium and low effort levels, with xhigh as the default, and preserved thinking now on by default.
- API price and limits
- $0.50 / $3 per million tokens instead of $0.60 / $3.60, with 1M tokens of context and 131K out instead of 262K and 65K.
Who is Qwen3.8-27B best for?
Qwen3.8-27B is best for developers who run coding agents on their own GPU or Mac, need one local model that also reads screenshots and video, or want a low-cost hosted fallback.
Coding agents on your own GPU
Alibaba scores it 61.7 on SWE-bench Pro, above Claude Opus 4.6 Max's 53.4, and its official FP8 weights take about 31 GB.
Long agent runs
Reasoning from earlier turns stays in context by default, which Qwen says keeps decisions consistent across multi-step tasks.
Screens, apps and documents
It scores 84.3 on OSWorld-Verified and 81.9 on AndroidWorld in Alibaba's table, so one local model can read screenshots and drive desktop or phone apps.
A low-cost hosted fallback
On Qwen Cloud it costs $0.50 / $3 per million tokens with a 1M-token context, less than Qwen3.6-27B's $0.60 / $3.60.
What's included with Qwen3.8-27B?
Adjustable reasoning effort, thinking kept across turns, image and video input, a 1M-token hosted context, official FP8 weights and built-in tools on Qwen Cloud.
- Reasoning effort levels
- xhigh by default, with medium and low to save time and tokens, and thinking can be switched off per request.
- Preserved thinking
- Reasoning from all earlier turns stays in the conversation by default, which Qwen recommends for agents.
- Image and video input
- The same weights read documents, charts, screenshots and hour-long videos.
- Long context
- 262,144 tokens natively, stretchable to 1M with YaRN in vLLM, SGLang or TokenSpeed; Qwen Cloud serves 1M by default.
- Official FP8 weights
- About 31 GB next to the 56 GB BF16 release, with multi-token prediction for faster decoding.
- Built-in tools on Qwen Cloud
- Web search, web extraction, code interpreter and image search are available through the hosted API.
How much does Qwen3.8-27B cost?
Qwen3.8-27B is free to download under Apache 2.0, and on Qwen Cloud it costs $0.50 per million input tokens and $3 per million output tokens, with 1M free tokens for new accounts.
- Apache 2.0 licence, on Hugging Face
- BF16 about 56 GB; official FP8 about 31 GB
- 262,144-token context, up to 1M with YaRN
- Cached input $0.10 per million tokens
- 1M-token context, up to 131K tokens out
- 1M free tokens for new accounts, for about 90 days
Qwen Cloud is Qwen's official API and lists prices in US dollars. Alibaba Cloud Model Studio charges the same $0.50 / $3 in its Singapore region and $0.424 / $1.696 in China (Beijing), and does not offer Qwen3.8-27B in its US (Virginia) region. Thinking tokens are billed as output, batch calls cost half, and the free quota needs payment details. Qwen Chat and the Token Plan do not include this model. Self-hosting costs only your own hardware. Full pricing on qwencloud.com. We checked these prices on .
How do readers rate Qwen3.8-27B?
No one has rated Qwen3.8-27B yet. Be the first.
Sign in to rate. One rating per account.
Qwen3.8-27B pros and cons
It is the strongest Qwen model most people can run at home and a clear step up from Qwen3.6-27B, but it overthinks by default and runs slowly on Macs and smaller GPUs.
Pros
- Many developers call it the first local model that does real work for them, from overnight agent runs to tracking down framework bugs, on one high-end GPU or Mac.
- Users moving from Qwen3.6-27B report it fixing bugs and finishing tasks that 3.6 could not, in line with Alibaba's jump from 53.5 to 61.7 on SWE-bench Pro.
- At 4 bits it fits a 24 GB card, where users report about 30 to 60 tokens a second on cards such as the RTX 4090 and RX 7900 XTX.
- One Apache 2.0 model covers code, screenshots and documents, which users running mixed local workloads value over text-only rivals.
Cons
- It overthinks at its default xhigh effort: one widely shared test drawing took 21 minutes on a MacBook Pro, and other users report tasks running for hours and 5 to 10 times the reasoning tokens of similar models.
- As a dense model it is slow on Macs, laptops and 12 to 16 GB cards, with users reporting from about 2 to 17 tokens a second there, far below mixture-of-experts models like Qwen3.6-35B-A3B.
- Several users find it weaker than larger models on general knowledge and open-ended chat, and some still prefer Qwen3.6-27B for conversation.
- The launch chat templates needed community fixes for reasoning levels and tool calls, and Qwen's own Token Plan and Qwen Chat do not offer the model.
Is Qwen3.8-27B open source?
Yes. Qwen3.8-27B's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Qwen3.8-27B alternatives?
Qwen3.8-27B is an open source alternative to Holo4.
Holo4FreemiumH Company's open computer use models that click, code and call tools.
Gemma 4Open SourceChoose it when speed on modest hardware matters more than peak coding scores: its 26B A4B model works only 4B parameters per token and loads in about 14.4 GB at 4-bit, and its smaller sizes take audio.
Qwen3.8-MaxFreeChoose it when you do not need to self-host and want Alibaba's strongest model: it scores 67.7 on SWE-bench Pro, is free in Qwen Chat and costs $2 / $6 per million tokens on Qwen Cloud.
DeepSeek V4.1 FlashOpen SourceChoose it when you want a cheaper hosted API with MIT-licensed weights: it costs $0.30 / $1.20 per million tokens at peak and half that off-peak, with a 1M-token context.
Questions about Qwen3.8-27B
What is Qwen3.8-27B?
Qwen3.8-27B is Alibaba's free, Apache 2.0 open-weight 27B model that reads text, images and video, built for developers who run coding agents on their own hardware or want a low-cost API.
Who is Qwen3.8-27B for?
Qwen3.8-27B is best for developers who run coding agents on their own GPU or Mac, need one local model that also reads screenshots and video, or want a low-cost hosted fallback.
Is Qwen3.8-27B free?
Yes. Qwen3.8-27B has a free Open weights plan: apache 2.0 licence, on Hugging Face.
How much memory does Qwen3.8-27B need?
Qwen's BF16 weights take about 56 GB and its official FP8 build about 31 GB, before extra memory for context. Qwen serves it with vLLM, SGLang or TokenSpeed, and smaller community quantisations for llama.cpp, LM Studio and Ollama are listed on Hugging Face.
Is Qwen3.8-27B open source?
Yes. Qwen3.8-27B's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Qwen3.8-27B alternatives?
The closest alternatives on Recaply are Holo4, Gemma 4, Qwen3.8-Max and DeepSeek V4.1 Flash.
Where can I try Qwen3.8-27B?
On its official website, huggingface.co.
Can I use Qwen3.8-27B commercially?
Yes. Qwen3.8-27B's weights are released under the Apache 2.0 licence, which allows commercial use, changes and redistribution as long as you keep the licence and its notices with the model. Qwen's official FP8 build uses the same licence.
How do I make Qwen3.8-27B think less?
Qwen3.8-27B defaults to xhigh reasoning effort. Set reasoning_effort to medium or low, or turn thinking off by setting enable_thinking to false. Qwen warns that lower effort can cause more failed steps and retries in long agent tasks, so test it on your own work.
Is Qwen3.8-27B better than Qwen3.7-Plus?
On coding and computer use, yes by Alibaba's numbers: 61.7 against 57.6 on SWE-bench Pro and 84.3 against 73.3 on OSWorld-Verified. Qwen3.7-Plus scores a little higher on GPQA Diamond and HLE, and costs less per output token on Qwen Cloud, at $1.60 against $3.
Our verdict
Qwen3.8-27B is the open model to pick if you want a capable coding agent on your own hardware.
Alibaba's tests put it well ahead of Qwen3.6-27B and above its own Qwen3.7-Plus on coding, and users confirm the jump, though it gets there by thinking at length.
- Skip it for
- People on Macs, laptops or 16 GB cards who need quick replies, and anyone who wants it inside Qwen Chat or a flat monthly plan.
- Check first
- Try medium reasoning effort on your own tasks and see whether the quality holds.
Qwen3.6-27B
Qwen3-Coder-Next
MiniMax M2.1
Rnj-1
Mellum2
Ornith-1.0