What it is
Qwen3.8-27B is the mid-sized open model in Alibaba's Qwen3.8 generation, published on Hugging Face in August 2026. It is a dense model, so all 27 billion parameters work on every token, and it uses the same hybrid design as Qwen3.5: three layers of fast linear attention (Gated DeltaNet) for every layer of standard attention. With its vision encoder the BF16 checkpoint is about 56 GB.
It reads text, images and video and writes text. The native context is 262,144 tokens, extendable to about 1 million with YaRN scaling. Thinking is on by default and can be switched off per request, and a reasoning_effort setting (xhigh by default, or medium or low) trades depth for speed. An official FP8 version (about 31 GB) is also published.
The weights use the standard Apache 2.0 licence, unlike Qwen3.8-Max and Qwen3.8-Flash-Next, which carry Qwen's own terms. Qwen Cloud lists a hosted version with a 1M-token context at $0.50 per million input tokens and $3.00 per million output tokens ($0.10 for cached input), and OpenRouter routes it through more than 20 providers.
Our review
Qwen3.8-27B is now the strongest case for running a serious model on hardware you own. On Alibaba's table it scores 61.7 on SWE-bench Pro, ahead of the 57.6 it reports for its own Qwen3.7-Plus and the 53.4 Anthropic published for Claude Opus 4.6 Max, and 84.3 on OSWorld-Verified, a computer-use test, against Opus's 72.7. The jump from its predecessor is steep: 42.2 against 13.3 for Qwen3.6-27B on DeepSWE 1.1.
Read those figures with care. Alibaba ran every test itself, mostly inside the Claude Code harness, and fixed tasks it judged faulty in SWE-bench Pro and answer keys in MathVision and CharXiv, so the scores are not directly comparable with results on the original benchmarks, as DataNorth points out. Two of the headline tests, QwenSWEBench and CoWorkBench, are in-house. Opus 4.6 Max still leads where breadth matters: 78.2 against 73.0 on Terminal Bench 2.1 and 40.0 against 30.8 on Humanity's Last Exam.
The memory maths is what makes it practical. The Unsloth 4-bit build is 16.5 GB and the 8-bit 29 GB, so one 24 GB or 32 GB card runs it, and the official FP8 checkpoint at 31 GB suits a single 48 GB or 80 GB server GPU. The cost is speed: a dense 27B model computes far more per token than a mixture-of-experts model such as Qwen3.6-35B-A3B, which uses 3 billion parameters at a time and has no Qwen3.8 counterpart on Hugging Face.
Hosted, it is close to free. OpenRouter lists providers from about $0.04 per million input tokens and $1.25 per million output, less than a tenth of Qwen Cloud.s own input price. The catch is the card's own warning that lower reasoning effort causes more failed attempts in agent loops, so the cheap, fast setting can end up using more tokens than the default.
Where it earns its place
A developer with a 24 GB or 32 GB workstation GPU
Run a local coding agent on a private repository
Load the 16.5 GB 4-bit GGUF in llama.cpp or a similar runtime and point a Claude Code-style harness at it; Alibaba's coding scores were measured in that harness. No per-token bill, but expect fewer tokens a second than a 3B-active MoE model on the same card.
A platform team serving an internal assistant to 300 staff on one H100
Host a permissively licensed model that also reads screenshots and PDFs
Serve the official 31 GB FP8 checkpoint with vLLM or SGLang, leaving room on an 80 GB card for the KV cache. Apache 2.0 means no revenue or user caps to track, unlike Qwen3.8-Flash-Next.
A SaaS company adding document Q&A for 5,000 customers
Answer questions over uploaded files without paying frontier prices
Use a hosted endpoint: at Qwen Cloud's $0.50/$3.00, 1 million requests at 4K tokens in and 500 out costs about $3,500 a month; DeepInfra on OpenRouter at $0.15/$1.875 brings that to about $1,540.
Before you commit
- Setup
- Weights on Hugging Face and ModelScope; the card recommends current SGLang, vLLM or TokenSpeed. Community GGUF builds run in llama.cpp. Thinking is on by default; set enable_thinking false or lower reasoning_effort per request.
- Integrations
- OpenAI-compatible serving in vLLM and SGLang; Qwen Cloud supports function calling, structured output, batches, caching and fine-tuning through OpenAI-style and Anthropic-style APIs.
- Data and privacy
- Apache 2.0: commercial use with no revenue caps or acceptable-use annex. Qwen Cloud's listing publishes no retention or training terms, and its sample endpoint is in Singapore (ap-southeast-1).
- Cost at scale
- Self-hosted: 4-bit 16.5 GB, 8-bit 29 GB, FP8 31 GB, BF16 about 56 GB of weights before KV cache. Hosted: Qwen Cloud $0.50/$3.00 per 1M tokens; OpenRouter providers range from about $0.04/$1.25 to $0.45/$3.20.
- Limits
- 262K native context, about 1M with YaRN, which the card says may degrade short inputs. Recommended output budget 131K tokens plus up to 262K of reasoning. Qwen Cloud caps throughput at 5M tokens a minute.
What's good
- Apache 2.0 licence with no usage conditions
- Vendor-reported SWE-bench Pro and OSWorld scores above Opus 4.6 Max
- 4-bit build of 16.5 GB fits one 24 GB GPU
- Image and video input; thinking adjustable per request
What's not
- All benchmarks run by Alibaba, some on modified or in-house tests
- Dense design is slower per token than small mixture-of-experts rivals
- Trails Opus 4.6 Max on Terminal Bench 2.1 and knowledge tests
- YaRN scaling past 262K can hurt short-prompt quality
Best for
Developers who want the strongest permissively licensed model that fits on one high-end GPU, especially for coding and computer-use agents.
Alternatives
Qwen3.8-27B: common questions
- Is Qwen3.8-27B worth it?
- A 27B model with an Apache licence that, on Alibaba's own numbers, edges past Claude Opus 4.6 Max at coding and fits on a single 24 GB card at 4-bit. We rate it 4.5 out of 5.
- How much does Qwen3.8-27B cost?
- Free weights under Apache 2.0. Qwen Cloud: $0.50 input / $3.00 output per 1M tokens ($0.10 cached). Third-party hosts on OpenRouter from about $0.04 / $1.25.
- What are the downsides of Qwen3.8-27B?
- All benchmarks run by Alibaba, some on modified or in-house tests; dense design is slower per token than small mixture-of-experts rivals; trails Opus 4.6 Max on Terminal Bench 2.1 and knowledge tests; yaRN scaling past 262K can hurt short-prompt quality.
- Who is Qwen3.8-27B best for?
- Developers who want the strongest permissively licensed model that fits on one high-end GPU, especially for coding and computer-use agents.
- How much does Qwen3.8-27B cost for a team or in production?
- Self-hosted: 4-bit 16.5 GB, 8-bit 29 GB, FP8 31 GB, BF16 about 56 GB of weights before KV cache. Hosted: Qwen Cloud $0.50/$3.00 per 1M tokens; OpenRouter providers range from about $0.04/$1.25 to $0.45/$3.20.
- Is Qwen3.8-27B safe for company data?
- Apache 2.0: commercial use with no revenue caps or acceptable-use annex. Qwen Cloud's listing publishes no retention or training terms, and its sample endpoint is in Singapore (ap-southeast-1).
- What are the best Qwen3.8-27B alternatives?
- The closest alternatives we have reviewed are Gemma 4, Muse Glimmer, Qwen3.8-Flash-Next and Mistral Small 4.
Answers checked Oct 22, 2026.
Make Qwen3.8-27B? Show this rating on your site
