Qwen3.8-Flash
Overview
Qwen3.8-Flash is a hosted multimodal model from Alibaba's Qwen team, announced on August 26, 2026. It is the production version based on Qwen3.8-Flash-Next, adding a default 1M-token context window and hosted tools for cost-sensitive coding, agentic, and visual knowledge-work applications.
Qwen3.8-Flash
Free access is available in phases across B.AI API and Chat:
- API: Qwen3.8-Flash API usage is currently billed at
0 Credits. No input, cache write, cache read, or output token fees apply. - Chat: Free access begins when Qwen3.8-Flash becomes available in B.AI Chat. The availability date is subject to the actual model listing. Once available, Chat usage is billed at
0 Credits.
After the offer ends, Qwen3.8-Flash will return to the prices shown on this page.
Key Features
- Production Flash-Next Lineage: Qwen identifies
qwen3.8-flashas the production version based on Qwen3.8-Flash-Next. The related open-weight architecture uses Gated DeltaNet, Qwen Sparse Attention, Gated Residual, and N-gram Embedding; QwenCloud does not separately publish the production model's parameter count. - Native Multimodal Input: Accepts text, images, and video and produces text output, supporting visual coding, document analysis, charts, and long-video understanding.
- 1M-Token Hosted Context: Supports up to 991K input tokens without thinking, 983K input tokens with thinking, and 131K output tokens in either mode.
- Thinking and Agent Controls: QwenCloud documents thinking as enabled by default for the Qwen3.8 series, exposes the
enable_thinkingcontrol, and lists a maximum reasoning budget of 262K tokens. - Agent-Oriented API Features: Supports prefix completion, function calling, context caching, structured output, Batch API processing, fine-tuning, and built-in tools through QwenCloud's Responses API.
- Published Flash-Next Evaluation: The related open-weight foundation reports 62.5 on SWE-bench Pro, 58.7 on DeepSWE 1.1, 73.9 on CoWorkBench, and 73.5 on Toolathlon Verified. Qwen has not published a separate benchmark table for the hosted production endpoint.
Best Use Cases
- Cost-Sensitive Coding Agents: Repository analysis, code generation, debugging, and tool-driven development where low token prices and high account-level rate limits matter.
- Long-Context Knowledge Work: Reviewing large document sets, codebases, conversation histories, and research materials within a 1M-token hosted context.
- Multimodal Analysis: Understanding screenshots, charts, scanned documents, interfaces, and video together with text instructions.
- Structured Agent Workflows: Applications that combine function calling, JSON structured output, code execution, search, extraction, and cached shared prompts.
- Asynchronous Bulk Processing: Classification, extraction, evaluation, and dataset processing through the Batch API at half the real-time input and output rates.
Capabilities and Limitations
| Capability | Description |
|---|---|
| Reasoning | Thinking is enabled by default for the Qwen3.8 series and can be controlled with enable_thinking. QwenCloud lists a 262K-token maximum reasoning budget but does not publish a model-specific reasoning-effort mapping on the model page. |
| Creative Writing | Supports general, long-form, and structured text generation. |
| Coding | The related Qwen3.8-Flash-Next evaluation reports SWE-bench Pro: 62.5, DeepSWE 1.1: 58.7, SWE-bench Multilingual: 81.0, and NL2Repo-Bench: 48.1. These are not hosted-endpoint SLA results. |
| Multimodal | Accepts text, image, and video input and produces text output. |
| Context Window | 1M tokens. |
| Maximum Input | 991K tokens in non-thinking mode and 983K tokens in thinking mode. |
| Max Output | 131K tokens in both thinking and non-thinking modes. |
| Tool Use | Supports function calling, structured output, prefix completion, caching, and Batch API processing. Responses API tools include code_interpreter, i2i_search, t2i_search, web_extractor, and web_search. |
| Multilingual | The related Flash-Next evaluation includes multilingual reasoning and coding benchmarks. |
Known Limitations
qwen3.8-flashis the hosted production model, whileQwen/Qwen3.8-Flash-Nextis the related open-weight architecture release. Parameter counts, self-hosting behavior, and Flash-Next benchmark results should not be treated as hosted-endpoint guarantees.- QwenCloud does not publish a model-specific knowledge cutoff or complete supported-language list.
- Thinking tokens are billed at the output-token rate and consume context. Applications should enable thinking according to task needs rather than assuming that a larger reasoning budget is always more efficient.
Pricing
| Model | Input (Credits/Token) | Cache Write (Credits/Token) | Cache Read (Credits/Token) | Output (Credits/Token) | Web Search (Credits/Use) |
|---|---|---|---|---|---|
| Qwen3.8-Flash | 0.16 | 0.16 | 0.016 | 0.47 | - |
Explicit cache creation costs 0.20 Credits/Token. Both explicit and implicit cache hits cost 0.016 Credits/Token.
Pricing note
Prices shown in the documentation are B.AI standard reference prices for base billing purposes. B.AI may provide lower actual usage costs through limited-time offers, top-up bonuses, and account benefits. Specific prices, bonus Credits, account benefits, and final billing are subject to the platform display and billing records.