Qwen Models: Qwen3.8 Sizes, Licences and Prices

Qwen is Alibaba's model family, and Qwen3.8 is the current open generation as of 31 October 2026. The one most people should try is Qwen3.8-27B: Apache-2.0, dense, 262,144 tokens of native context and image and video input. The flagship, Qwen3.8-2.4T-A95B, has 2.4 trillion parameters and a custom licence. The hosted versions are Qwen3.8-Max, 27B and Flash.

The Qwen3.8 lineup

All facts below come from the Hugging Face cards and API metadata, plus Alibaba Cloud's pricing page, read on 31 October 2026.

There are also FP8 builds of the 27B and Flash-Next models on Hugging Face, which roughly halve memory against BF16.

What the names mean

Qwen publishes weights and a hosted product that share a family but differ in features. The 2.4T card says Qwen3.8-Max is the official version based on those weights, with extras such as vision input, non-thinking support, a default 1M context and built-in tools. The 27B card says a hosted version with 1M default context and built-in tools is coming. So the downloadable model is the base, and the hosted one is the product. If you self-host the 2.4T model you do not get Max's vision input.

Licences

Qwen3.8-27B is Apache-2.0. Nothing to report, nothing to renegotiate.

Qwen3.8-2.4T-A95B carries the Qwen3.8-Max License, which we read. It starts as a broad permission grant and then adds two conditions.

• Products with more than 100 million monthly active users built on the model or derivatives have a display condition (the file's text is cut in our extraction after that clause).
• A licensee that runs a Model-as-a-Service or "AI Work Assistant" business, with group revenue above 50 million US dollars over twelve months, has an additional obligation. The licence names coding and office-productivity assistants such as Qoder and QwenWork as examples of an AI Work Assistant.

What the licensee must do once those lines are crossed is not in the text we captured:. The AI Work Assistant wording is unusual. A company that sells a coding agent powered by the weights should read it with a lawyer.

Qwen3.8-Flash-Next is tagged qwen-community-1.0. We did not read it:.

Hosted prices

Alibaba Cloud Model Studio lists per-million-token prices. The figures below are the international rates on its pricing page, checked 31 October 2026. Alibaba says the page shows standard prices only and that promotions appear in the console.

Model Studio also lists a Beijing region with separate lower rates, for example qwen3.8-2.4t-a95b at $1.65 input and $4.951 output. Billing details we saw: batch calls at 50 percent for supported models, cache hits at 10 percent of input price, and explicit cache creation at 125 percent. Which of these apply to which 3.8 model is not clear from the page:. One more caution: the page is a long table and our fetch flattened it, so confirm the price on the console before you commit a budget.

Set these beside the closed models. On our LLM API price comparison, Sonnet 5.5 is $2 and $10 and Opus 5.5 is $4 and $20. Qwen3.8-Max at $2 and $6 undercuts Sonnet on output, and 27B at $0.50 and $3 undercuts it much more.

Published scores

The following are vendor-reported on Qwen's own cards. Qwen ran its own models and, as the card notes explain, reported the best published score across harnesses for others.

Qwen3.8-Max (the 2.4T weights, hosted)

Source: Qwen3.8-2.4T-A95B model card. Terminal Bench 2.1 for Qwen was run with Claude Code (average of 10 runs, 5-hour timeout); the other models' scores are the best published across harnesses. A note on the card also says Fable 5 results may involve fallbacks.

Qwen3.8-27B

Source: Qwen3.8-27B model card, vendor-reported. On that card Claude Opus 4.6 Max is shown at 78.2 on Terminal Bench 2.1 and 53.4 on SWE-bench Pro. The 27B model beating an older Opus on SWE-bench Pro is Qwen's claim, with problem tasks "corrected" by Qwen, which is a different benchmark than the public version. Treat it as a pointer.

What Z.ai's card says about Qwen3.8-Max

Z.ai's GLM-5.3 card lists Qwen3.8-Max at 86.6 on Terminal Bench 2.1, 56.6 on DeepSWE and 55.9 on NL2Repo, matching Qwen's own table, and 78.5 on CyberGym. It is a useful cross-check because a rival repeated the numbers. See our best open source LLM for coding page for the cross-lab table.

Running Qwen3.8-27B yourself

The card recommends SGLang, vLLM or TokenSpeed for production, and works with Hugging Face Transformers. Our arithmetic puts the weights at about 56 GB in BF16, so the realistic options are:

The card's guidance matters in practice:

• Thinking is on by default. Turn it off per request for fast answers. With thinking on, use temperature 1.0, topp 0.95 and topk 20. For non-thinking, the card recommends temperature 0.7, top_p 0.80 and a presence penalty of 1.5.
• Preserved thinking is on by default, which keeps reasoning from earlier turns. That helps agents stay consistent but uses more context.
• For anything past 262,144 tokens, set YaRN rope scaling in the model configuration. The card shows the exact override for vLLM, SGLang and TokenSpeed, and advises changing it only when you need long inputs, because a fixed scaling factor can slightly hurt short-text quality.

For a serving-tool comparison read Ollama vs vLLM. Ollama's catalogue names for Qwen3.8 were not confirmed:.

The 2.4T model is a different class of job. At about 2.4 trillion parameters the repository's weight files total roughly 4.9 TB according to the Hugging Face API, so it needs a multi-node GPU deployment. Rent GPUs from Runpod, Modal or Lambda only if you have a strong reason not to use the API.

First run: serve the 27B with vLLM

The card gives full recipes; the shape of a first run is short. Install a recent vLLM, point it at the Hugging Face repository and set the context you actually need rather than the maximum.

vllm serve Qwen/Qwen3.8-27B --max-model-len 262144

Treat that line as a sketch, not the card's exact command:. Once the server is up, any OpenAI-compatible client can call it. Send a thinking request first and a non-thinking one second, and compare latency. With thinking on, the model writes reasoning inside think tags before the answer, so budget a few thousand extra output tokens per request for hard prompts. For an agent loop, leave preserved thinking on, since the card says it helps decisions stay consistent between turns.

Cost arithmetic at agent scale

Hosted price only means something against a workload. Take an agent run that reads 3 million input tokens and writes 300,000 output tokens, which is plausible for a multi-hour repository task. At Alibaba's international rates that is $6.00 for input plus $1.80 for output on qwen3.8-max ($7.80), $1.50 plus $0.90 on qwen3.8-27b ($2.40), and $0.45 plus about $0.14 on qwen3.8-flash (about $0.59). Cache hits and batch discounts would lower those, and long thinking traces would raise the output side. The point is the ratio: the same run is about three times cheaper on 27B than on Max and about thirteen times cheaper on Flash, before accounting for how many runs fail. Failures matter more than list price, so measure cost per merged change on your own repositories.

Choosing between 27B and Flash-Next

Both are open, both accept images, and both extend to 1M tokens. They differ in what they trade.

A mixture-of-experts model like Flash-Next is quick per token but still needs every expert in memory. That makes the dense 27B the better fit for one GPU despite its higher per-token compute. Flash-Next suits a server with several GPUs where you want high throughput. Its card describes a hybrid attention design aimed at cutting long-context latency, which Qwen says is a design goal; we have no measured speed figures to report:.

Where Qwen is served

Alibaba Cloud Model Studio is the maker's own API, with international and Beijing regions. Qwen3.8-27B's card links Qwen Cloud (qwencloud.com) as the official managed service. Third-party hosts such as Hugging Face Inference, Replicate and Cloudflare Workers AI serve various Qwen models, but we did not confirm which Qwen3.8 builds each carries:.

Which coding agents can use Qwen

Any agent with a custom OpenAI-compatible endpoint can call a Qwen model served by vLLM or SGLang, or Alibaba's API. Qwen's own benchmark notes say Qwen3.8-Max scored best inside the Claude Code harness, and Z.ai ran Qwen3.8-Max in Claude Code for its ExploitGym tests, so Claude Code is a workable front end. Alibaba's licence text mentions Qoder as its own AI Work Assistant product. For third-party options see Claude Code, OpenCode and Cline. Anthropic-format endpoint support on Alibaba Cloud was not confirmed:.

Who should pick Qwen

Limits to know about

• Text-only flagship. The downloadable 2.4T model has no vision input. For images at scale use 27B or Flash-Next.
• Thinking cannot be disabled on the flagship. Every reply starts with reasoning, so latency and token use are higher than a non-thinking model.
• Native 256K, extended 1M. The long-context mode relies on YaRN and costs memory.
• Licence differences inside one family. Three models, three licences. Do not assume the 27B terms cover the flagship.

Our take

Qwen3.8-27B is the easiest recommendation on this page: permissive, mid-sized, multimodal and well documented. For a hosted flagship, Qwen3.8-Max is priced below Claude Sonnet 5.5 on output, though we have no independent comparison of quality. The part to watch is the licence on the largest weights, which includes a clause aimed at coding and productivity assistants. See open source LLM for the full licence table and DeepSeek models for the closest MIT-licensed rival.

Product names are used descriptively. We are not affiliated with Alibaba.