What it is
Qwen3.8-Flash-Next is an open-weight mixture-of-experts model that Alibaba's Qwen team released in late August 2026 as an experimental preview of the architecture behind Qwen4. The language model has 125 billion parameters, of which 6 billion are active for each token, plus 51 billion in a new n-gram embedding table and 4 billion in a multi-token prediction module, about 180 billion on disk in all. Weights are published in BF16 and FP8.
Every fourth layer uses Qwen Sparse Attention, which picks small blocks of earlier text to attend to rather than individual tokens, cutting latency on long inputs. The n-gram embeddings add capacity that, Qwen says, is cheaper to run and easier to offload from the GPU than extra experts. It reads text, images and video, thinks by default, and has a native context of 262,144 tokens, extendable to 1 million.
The weights use the Qwen Community License 1.0, not Apache 2.0. Commercial use is allowed, but products above 100 million monthly users or US$20 million in monthly revenue must display the model name, and any company selling model hosting or an AI coding or office assistant needs a separate licence from Qwen, whatever its size. The hosted version, Qwen3.8-Flash, costs $0.15 per million input tokens and $0.47 per million output tokens on Qwen Cloud, with a 1M-token context and up to 131K output tokens.
Our review
Flash-Next edges past Qwen's dense 27B on coding and clearly beats it on agent work while activating a fraction of the parameters, but it comes with a stricter licence than Qwen's 2.4-trillion-parameter flagship. Qwen's table puts it at 62.5 on SWE-bench Pro against 61.7 for Qwen3.8-27B, and at 58.7 on DeepSWE 1.1 against 42.2 for the 27B and 54.4 for DeepSeek-V4-Flash-0731, with 6 billion active parameters to DeepSeek's 13 billion. The widest gap is on JobBench, 55.7 to 33.4.
It does not win everywhere, and the scores are Qwen's own. DeepSeek-V4-Flash-0731 leads on NL2Repo (54.2 against 48.1) and on the Agents' Last Exam pass rate (25.2 against 24.3). Qwen corrected tasks in SWE-bench Pro before re-running every model, so its figures do not line up with public leaderboards; NVIDIA's Nemotron 3.5 card reports 60.37 for Flash-Next on SWE-bench Pro.
The licence is the catch. Qwen Community License 1.0 requires a separate licence before any commercial use by a company, or its affiliates, that runs a model-as-a-service business or sells a standalone AI coding or office assistant, with no size threshold. The Qwen3.8-Max licence applies the same rule only above $50 million in yearly revenue, so a five-person start-up building a coding assistant can ship on the flagship's weights but not on these. Internal use that exposes nothing to third parties is exempt, and Qwen3.8-27B's Apache 2.0 and GLM-5.3-Flash's MIT carry no such terms.
Small active size does not mean a small server. All 180 billion parameters, including the 51 billion in the n-gram table, have to be loaded: about 360 GB in BF16 or 180 GB with Qwen's FP8 weights by our arithmetic, before any long-context cache. The hosted Qwen3.8-Flash at $0.15 input and $0.47 output per million tokens avoids the hardware, and API customers are governed by Qwen Cloud's terms rather than the weights licence.
Where it earns its place
A platform team at a 300-engineer company running an internal coding agent
Self-host something stronger than Qwen3.8-27B without sending code outside
Serve the FP8 weights on vLLM or SGLang; at about 180 GB of weights that means a multi-GPU server such as four 80 GB H100s, by our arithmetic. Internal use is exempt from the hosting clause as long as nothing is exposed to outside users.
A founder building a paid AI coding assistant
Choose an open model to fine-tune and ship
This licence requires a separate agreement with Qwen before any commercial launch, whatever the company's revenue. Qwen3.8-27B (Apache 2.0) or GLM-5.3-Flash (MIT) avoid the negotiation, and Qwen3.8-Max's licence only applies the rule above $50 million a year.
A developer running a screenshot and document triage agent
Process images and long inputs cheaply through an API
Call hosted Qwen3.8-Flash with thinking off. 1 billion input and 100 million output tokens a month cost about $197, less where repeated prefixes hit the cache at $0.016 per million; throughput is capped at 2 million tokens a minute.
Before you commit
- Setup
- BF16 and FP8 weights on Hugging Face and ModelScope, served with Transformers, vLLM, SGLang or TokenSpeed. Going beyond 262K tokens needs a YaRN override in the serving command.
- Integrations
- Qwen Cloud serves Qwen3.8-Flash through OpenAI-compatible Chat Completions and Responses, an Anthropic-compatible API and DashScope, with function calling, structured outputs and built-in web search, code interpreter and PDF parsing.
- Data and privacy
- Qwen Community License 1.0: commercial use allowed; model name must be shown above 100 million monthly users or $20 million monthly revenue; separate licence for model-as-a-service and AI coding or office assistant businesses at any size. Qwen Cloud's retention and training terms are not stated on the model page.
- Cost at scale
- Hosted: $0.15 input and $0.47 output per million tokens and $0.016 for cache hits; Batch is offered with no published discount. Self-hosted: roughly 360 GB of BF16 or 180 GB of FP8 weights, so a multi-GPU H100 or H200 server before KV cache.
- Limits
- 262,144 tokens native, 1 million with YaRN; hosted output up to 131K tokens at 2 million tokens a minute. Thinking is on by default and earlier reasoning stays in context unless preserve_thinking is turned off, which adds tokens on long agent runs.
What's good
- Only 6 billion active parameters per token
- Beats Qwen3.8-27B on every coding and agent test in Qwen's table
- Image and video input; 262K native context, 1M extended
- Hosted version at $0.15 / $0.47 per million tokens
What's not
- Separate licence needed for any hosting or AI-assistant business, at any size
- About 180 billion parameters to hold in memory
- Experimental preview of a new architecture
- Benchmarks are Qwen's own, on corrected test sets
Best for
Teams running internal coding or agent workloads that outgrow a 27B model, and developers who want a very cheap hosted multimodal API.
Alternatives
Qwen3.8-Flash-Next: common questions
- Is Qwen3.8-Flash-Next worth it?
- Strong agent and coding scores for 6 billion active parameters, under a licence that is stricter than the one on Qwen's own flagship. We rate it 4.2 out of 5.
- How much does Qwen3.8-Flash-Next cost?
- Open weights under the Qwen Community License 1.0 (separate licence for hosting and AI-assistant businesses). Qwen3.8-Flash API: $0.15 input / $0.47 output per 1M tokens; cache hits $0.016.
- What are the downsides of Qwen3.8-Flash-Next?
- Separate licence needed for any hosting or AI-assistant business, at any size; about 180 billion parameters to hold in memory; experimental preview of a new architecture; benchmarks are Qwen's own, on corrected test sets.
- Who is Qwen3.8-Flash-Next best for?
- Teams running internal coding or agent workloads that outgrow a 27B model, and developers who want a very cheap hosted multimodal API.
- How much does Qwen3.8-Flash-Next cost for a team or in production?
- Hosted: $0.15 input and $0.47 output per million tokens and $0.016 for cache hits; Batch is offered with no published discount. Self-hosted: roughly 360 GB of BF16 or 180 GB of FP8 weights, so a multi-GPU H100 or H200 server before KV cache.
- Is Qwen3.8-Flash-Next safe for company data?
- Qwen Community License 1.0: commercial use allowed; model name must be shown above 100 million monthly users or $20 million monthly revenue; separate licence for model-as-a-service and AI coding or office assistant businesses at any size. Qwen Cloud's retention and training terms are not stated on the model page.
- What are the best Qwen3.8-Flash-Next alternatives?
- The closest alternatives we have reviewed are GLM-5.3-Flash, DeepSeek-V4.1-Flash, Qwen3.8-27B and Nemotron 3.5 Super VL.
Answers checked Oct 22, 2026.
Make Qwen3.8-Flash-Next? Show this rating on your site
