Your privacy choices

Allow optional cookies for referral attribution, visit analytics, and Google Ads purchase measurement.

Back to blog
العربيةDeutschEnglishEspañolFrançais日本語한국어LatinaBahasa MelayuPortuguês (Brasil)РусскийไทยTiếng Việt繁體中文中文

Qwen3.8-Flash API Pricing & Coding Guide 2026: 1M Multimodal

Qwen3.8-FlashQwen APIQwen FlashCoding Agent1M context

On August 27, 2026, the day after Qwen3.8-Flash-Next became open source, we prepared integration notes for the qwen3.8-flash route newly opened on this site. The easiest mistake is in the name itself: the open weights and the production API do not use the same model ID.

Qwen3.8-Flash-Next is an open multimodal MoE for previewing the Qwen4 architecture; qwen3.8-flash is the production API model offered by Alibaba Cloud Model Studio and QwenCloud, with 1M context by default, image/text/video input, Function Calling, and structured output. This site has opened the qwen3.8-flash route; the site multiplier and actual charges should be checked on the live pricing page and in the account ledger.

This article keeps the open model architecture, Alibaba Cloud’s official API capabilities, official direct-provider pricing, and this site’s route separate. Benchmarks published by the Qwen team are not independent tests by this site, and Alibaba Cloud regional prices cannot be converted directly into this site’s multiplier.

Qwen3.8-Flash key stats

Item Verified information
Official release August 26, 2026
Site / Model Studio API ID qwen3.8-flash
Open-weight ID Qwen/Qwen3.8-Flash-Next
Parameter structure 125B main model + 51B N-gram Embedding, 6B active parameters
Context Production API defaults to 1M; open version is native 262,144 and can extend to 1M with YaRN
Input / output modalities Image, text, and video input; text output
API capabilities Function Calling, structured output, Web Search, and context caching
Public positioning Coding Agent, office automation, visual understanding, and high-concurrency workflows
Site multiplier Not hard-coded in this article; check the live pricing page and actual ledger

Last verified on August 27, 2026. Before production launch, recheck the model list, Alibaba Cloud regional documentation, and usage fields in actual responses.

What is the relationship between Qwen3.8-Flash and Qwen3.8-Flash-Next?

The Qwen team first published Qwen3.8-Flash-Next weights so the community could study the architecture planned for Qwen4. The production version is served under the name Qwen3.8-Flash. They share a technical direction, but their calling methods are different.

Name Use Exact ID Context
Qwen3.8-Flash-Next Self-hosting, architecture research, and open evaluation Qwen/Qwen3.8-Flash-Next Native 262,144; YaRN can extend it to 1M
Qwen3.8-Flash Hosted production API qwen3.8-flash 1M by default
Qwen3.8-Max Hosted flagship model for higher capability qwen3.8-max 1M

When calling through this site or an Alibaba Cloud-compatible interface, do not put the Hugging Face repository name in the model parameter. Conversely, when self-hosting, do not enter only the hosted API ID and expect the framework to download the weights automatically.

What is new in the Qwen3.8-Flash-Next architecture?

Qwen3.8-Flash-Next combines Gated DeltaNet (GDN) and Qwen Sparse Attention (QSA) in a hybrid attention design. GDN compresses history into a fixed state, while QSA uses a lightweight indexer to select important context by micro-block, reducing attention compute and KV Cache access cost on long sequences.

It also adds Gated Residual (GR), expanding the residual stream into four branches controlled by dynamic gates for reads and writes. N-gram Embedding looks up local context and increases model capacity without adding much per-token matrix computation. Training uses an improved Muon Optimizer and a refitted Scaling Law.

The official scale is 125B for the main model, an additional 51B N-gram Embedding, and 6B parameters activated per token. The Qwen team says its training cost is about one-ninth of Qwen3.7-Plus. This is a vendor training comparison, not a measurement of this site’s inference cost.

What are the boundaries of 1M context and multimodal capability?

The Alibaba Cloud Qwen3.8-Flash model page lists a context length of 1,000,000 tokens: maximum input of 991,808 in regular mode, 983,616 in thinking mode, maximum output of 131,072, and maximum thinking-chain length of 262,144. Images, text, and video can be inputs, while output is text.

1M does not mean that every request should approach 1M tokens. System prompts, tool schemas, image/video encoding, message history, and reserved output all consume the window. Very long requests also add latency and cost. Production systems should first test 32K, 128K, 256K, and long-context tiers.

The official capability table also lists Function Calling, structured output, Web Search, and context caching. Batch is supported in the Beijing region and marked unsupported in Singapore; verify regional differences separately before deployment.

How should you read Qwen’s official coding and agent benchmarks?

The Qwen team compared Qwen3.8-Flash-Next with Qwen3.8-27B, Qwen3.7-Plus, and DeepSeek-V4-Flash-0731. Four entries are excerpted below:

Vendor-published evaluation Qwen3.7-Plus Qwen3.8-Flash-Next Task type
DeepSWE 1.1 16.5 58.7 Agentic coding
SWE-bench Multilingual 75.8 81.0 Multilingual software engineering
CoWorkBench 65.1 73.9 Long-horizon office tasks
Toolathlon Verified 50.6 73.5 Real tool use

These scores come from the Qwen team’s official release, not independent tests by this site. Model versions, inference budgets, tool environments, and scorers can materially affect results. In particular, do not treat the open Flash-Next scores as end-to-end performance for a hosted API under your client, region, and quota.

How should you separate Qwen3.8-Flash official pricing from this site’s pricing?

At launch, the Qwen release post stated that the QwenCloud production version had reference pricing of USD 0.16 per million input tokens and USD 0.47 per million output tokens; the same post said the API would open soon. The Alibaba Cloud Model Studio page later listed the qwen3.8-flash call ID and different CNY prices and capability tables for regions such as Beijing and Singapore.

Both figures use official direct-provider pricing, and either can change with region, caching, Batch, or promotions. This site does not convert the official USD price or Model Studio regional price directly into a gateway multiplier. The multiplier, caching, and actual charge for this site’s qwen3.8-flash route are governed by the live pricing page and the request ledger.

New users can open a small balance from the API purchase page; existing key holders can add balance on the top-up page. Fix the prompt and context length first, record input, output, and cached tokens, and then compare the real cost.

How to call the Qwen3.8-Flash API

This site provides an OpenAI-compatible Chat Completions endpoint. The exact model ID is qwen3.8-flash:

curl https://api.llm-token.cn/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.8-flash",
    "messages": [
      {
        "role": "user",
        "content": "Summarize the failing tests, then propose the smallest safe patch."
      }
    ]
  }'

Qwen3.8-Flash supports thinking and non-thinking modes, but compatibility layers may map enable_thinking, reasoning_effort, tool-streaming parameters, and built-in tools differently. Start with the smallest text request, then add thinking, images, video, tools, and long context one at a time. Do not send every extension parameter in the first request.

Which tasks fit Qwen3.8-Flash?

  • High-concurrency Coding Agents, code review, and test-failure analysis
  • Cross-language repositories and SWE-bench-style software engineering
  • Long-horizon office automation for email, spreadsheets, documents, and meeting material
  • Image, chart, and long-video understanding
  • Function Calling, structured output, and Web Search workflows
  • Long documents and large codebases that need 1M context but are cost-sensitive
  • Research and inference-framework adaptation for the self-hosted Qwen4 preview architecture

Complex architecture decisions, critical security audits, and high-value one-off tasks should still be compared with Qwen3.8-Max, GLM-5.3, Claude, or another stronger model. Flash’s value is cost efficiency and throughput, not replacing every evaluation with one model.

Qwen3.8-Flash or Qwen3.8-Max: which should you choose?

If the workload is large and needs multimodality, tool calls, and 1M context, start with a small Qwen3.8-Flash traffic slice. If the task values the upper bound of complex reasoning, fault tolerance, and critical output quality, move the same tests to Qwen3.8-Max.

Do not compare one answer only. At minimum, track task pass rate, time to first token, total duration, tool-call success rate, input/output tokens, cache hits, and rework count. A model with a low unit price may not be cheaper overall if it requires more retries.

Production evaluation checklist

  1. Use qwen3.8-flash for the hosted API and Qwen/Qwen3.8-Flash-Next for self-hosting.
  2. Fix the repository, prompt, tool versions, timeout, and acceptance command before comparing models.
  3. Validate thinking and non-thinking modes separately and record response-format differences.
  4. Test text, images, video, Function Calling, and structured output separately.
  5. Measure latency and ledger cost in 32K, 128K, 256K, and long-context tiers.
  6. Verify whether the chosen region supports Batch, caching, and the built-in tools you need.
  7. Set limits for tool calls, maximum output, total budget, and external operations.
  8. Keep a fallback route for capacity shortages, rate limits, or protocol incompatibility.

Key takeaways

  • qwen3.8-flash is the exact model ID for this site’s and Alibaba Cloud’s hosted API.
  • Qwen/Qwen3.8-Flash-Next is the open-weight name and should not be entered as the hosted API model parameter.
  • The hosted version defaults to 1M context and supports image, text, and video input plus Function Calling.
  • It has a 125B main model, 51B N-gram Embedding, and 6B active parameters per token.
  • Official benchmarks and regional pricing must be kept separate from this site’s route behavior, multiplier, and ledger.
  • Production selection should consider pass rate, latency, usage, tool success, and rework cost together.

Frequently asked questions

Is Qwen3.8-Flash officially live?

Yes. The Qwen team released the Flash-Next weights on August 26, 2026, and the Alibaba Cloud official model page lists the production call ID qwen3.8-flash. This site has also opened the same route.

What context length does Qwen3.8-Flash support?

The hosted production API supports 1M context by default. The open Flash-Next weights support 262,144 tokens natively and can extend to 1,000,000 tokens with YaRN.

Does Qwen3.8-Flash support images and video?

Yes. Alibaba Cloud’s official capability table lists image, text, and video input and text output, along with Function Calling and structured output.

Are Qwen3.8-Flash-Next and Qwen3.8-Flash the same ID?

No. The former is the open repository and self-hosted weight name; the latter is the model value for the production API.

Does Qwen3.8-Flash support a thinking mode?

It supports thinking and non-thinking modes. Different compatible interfaces may map extension parameters differently, so validate the smallest request before adding thinking settings.

Can Qwen3.8-Flash be used with Claude Code or Codex?

Alibaba Cloud describes it as compatible with OpenAI and Anthropic protocols and names Claude Code and Codex. Actual integration still needs tests for tool calls, streaming, timeouts, and usage returns.

What is the official price of Qwen3.8-Flash?

The QwenCloud launch post gave USD 0.16 per million input tokens and USD 0.47 per million output tokens; Alibaba Cloud Model Studio has separate regional CNY pricing. This site’s multiplier and ledger are independent, so check the live pricing page.

Where can I buy or top up Qwen3.8-Flash API access?

New users can visit the API purchase page; existing key holders can use the top-up page.

Primary sources