Qwen3.8-Flash Explained: Pricing, 1M Context, Benchmarks, and Flash-Next

Verified August 28, 2026. Qwen3.8-Flash is Alibaba’s new hosted multimodal “Flash” model, exposed in Model Studio as qwen3.8-flash. It accepts text, images, and video, returns text, and offers a one-million-token context window. [S04, S10] The part most launch summaries miss is the naming: Qwen3.8-Flash and Qwen3.8-Flash-Next are related but not interchangeable products. Flash-Next is the open-weight architecture preview pointing toward Qwen4; the hosted Flash endpoint is the production service. [S01–S06]

The short version

Alibaba documents up to 991,808 input tokens, 131,072 output tokens, and 1,000,000 total context for hosted Qwen3.8-Flash. [S04] Thinking is enabled by default. [S09] Function calling, structured output, prefix continuation, and context caching are supported. [S04, S10, S13]

The current original invocation price in Beijing and Frankfurt is CNY 0.8 per million input tokens and CNY 2.7 per million output tokens. Singapore is CNY 1.094/3.427. QwenCloud and OpenRouter currently expose a public USD reference around $0.16/M input and $0.47/M output. [S04, S06, S18, S19]

Flash-Next is the model whose architecture has been published: a 125B-parameter main network plus 51B of N-gram embeddings, with 6B parameters active per token. It uses the Qwen Community License 1.0, not Apache-2.0. [S02, S36]

One launch, two names, two different decisions

On August 26, Alibaba released the production Qwen3.8-Flash service and opened Qwen3.8-Flash-Next weights. Qwen calls Flash-Next an early preview of the architecture intended for Qwen4. [S01, S02, S05]

That distinction changes what can safely be claimed. The 125B-main-model figure is explicitly documented for Flash-Next. Alibaba’s hosted qwen3.8-flash page does not present that architecture number as a cloud-service specification. Likewise, the Community License applies to the open weights, while hosted API usage is governed by service terms.

Question	Hosted Qwen3.8-Flash	Qwen3.8-Flash-Next
What is it?	Production API model	Open-weight architecture preview
Identifier	qwen3.8-flash	Qwen/Qwen3.8-Flash-Next on HF
Context	1M tokens	262K native; Qwen describes YaRN extension to 1M
Local weights	Not implied by hosted model ID	Yes
Published architecture	Cloud page focuses on capabilities/limits	125B main + 51B N-gram, 6B active/token
License	Hosted service terms	Qwen Community License 1.0

A genuinely large context window

For the hosted model, Alibaba lists 1,000,000 tokens of context, with a 991,808-token maximum input in non-thinking mode or 983,616 in thinking mode. Output can reach 131,072 tokens, and the maximum thinking-chain length is 262,144. [S04]

A million-token limit is useful for repository-scale code analysis, long document collections, persistent agent traces, and video-plus-text workloads. It should not be read as a recommendation to fill the window on every request: long prompts still affect latency, spend, retrieval precision, and failure modes.

Pricing changed almost immediately after launch reporting

Reuters’ August 26 report quoted the launch price at CNY 1/M input and CNY 3/M output. Alibaba’s live Model Studio documentation checked on August 28 already lists CNY 0.8/2.7 for Beijing and Frankfurt. Singapore is CNY 1.094/3.427. [S04, S06, S24]

For a fast-moving API, the current regional pricing page should therefore override a launch-day news article when the two disagree. For the larger model, the separate Qwen3.8-Max pricing and benchmark guide keeps that comparison distinct.

Qwen3.8-Flash is not cheaper than 3.7 at every prompt length

The comparison becomes more interesting against Qwen3.7-Flash. In Beijing/Frankfurt, qwen3.7-flash costs CNY 0.2/0.8 per million tokens for prompts up to 32K. Against Qwen3.8-Flash’s flat CNY 0.8/2.7, the newer model is 4× more expensive on short input and 3.375× on output.

At 256K–1M, however, qwen3.7-flash moves to CNY 1.2/4.8. Qwen3.8-Flash is then 33.3% cheaper on input and 43.75% cheaper on output. [S06]

That makes the economics workload-dependent. A short classification service may still favor Qwen3.7-Flash on pure token cost; long-context coding and agent workloads have a much stronger cost case for 3.8.

Monthly API cost examples

Using the currently listed $0.16/M input and $0.47/M output public API reference: [S18, S19]

• Light: 10M input + 2M output = $2.54/month
• Medium: 100M input + 20M output = $25.40/month
• Heavy: 1B input + 200M output = $254/month

These are transparent token-only calculations: (input millions × $0.16) + (output millions × $0.47). Search, storage, provider routing, caching behavior, taxes, and future price changes can alter the bill.

Why Flash-Next is architecturally interesting

Flash-Next combines Gated DeltaNet (GDN) with Qwen Sparse Attention (QSA). SGLang’s day-zero engineering report describes a 48-layer model: 36 GDN layers and 12 QSA layers, arranged as three GDN layers followed by one QSA layer. The sparse-attention path indexes compressed micro-blocks and retrieves a bounded set of relevant positions rather than attending densely over the full sequence. [S02, S21]

The model also adds Gated Residual, which expands the residual stream into four branches, and a large N-gram Embedding memory. Qwen lists 51B N-gram embedding parameters on top of a 125B main model, while only 6B parameters are active per token. SGLang shows how that embedding table can be kept in host memory and asynchronously prefetched, reducing permanent GPU-weight pressure. [S02, S21]

Qwen says training cost was about one ninth of Qwen3.7-Plus while delivering stronger coding and office-task capability. Treat that as a vendor efficiency claim; the public sources reviewed here do not provide an independent reconstruction of Alibaba’s training bill. [S02, S24]

Benchmarks: useful evidence, not a single truth score

Qwen reports 62.5 on SWE-bench Pro, 81.0 on SWE-bench Multilingual, 73.9 on CoWorkBench, 91.7 on GPQA Diamond, and 91.9 on LiveCodeBench v6 for Flash-Next. These are strong vendor-reported results, but harness choice, agent scaffolding, sampling, and tool configuration matter. [S01]

Artificial Analysis offers a different kind of evidence. At retrieval time it gives Flash-Next an Intelligence Index of 56 and reports roughly 73.4 output tokens/s with a 2.98-second time to first token on the tested Alibaba endpoint. [S20]

The two sets should stay separate: Qwen’s task benchmarks describe capability under its evaluation setups, while Artificial Analysis measures its own standardized index and serving characteristics.

Hosted API features and regional differences

Alibaba supports OpenAI-compatible Chat Completions and Responses APIs and an Anthropic-compatible Messages API. The model supports tools/function calling and structured outputs; thinking is enabled by default. [S07–S09, S12, S17]

The deployment region affects more than data location. Frankfurt supports context caching but, according to the current model page, not the built-in web-search feature for Qwen3.8-Flash. Beijing supports batch inference, while Singapore and Frankfurt do not for this model; Alibaba’s batch docs also cap a batch request at 256K context. [S04, S14, S15]

Rate limits differ too: Alibaba currently lists 30,000 RPM/5M TPM for Beijing, 15,000 RPM/2M TPM for Singapore, and 30,000 RPM/5M TPM for Frankfurt. [S16]

Local use: technically supported, not a small-model experience

For local inference, look at Flash-Next rather than the cloud model name. Qwen lists Transformers, llama.cpp, MLX/mlx-vlm, SGLang, vLLM, and other routes. Apple Silicon support is explicitly mentioned through MLX for text and vision. [S02] If you want to start with a smaller local checkpoint, the Qwen3.8 27B guide is a more realistic baseline.

Memory is the hard constraint. Unsloth says its extreme 1-bit Dynamic GGUF can run with about 75GB of RAM or unified memory. That is a third-party quantization claim, not an official minimum, but it illustrates the scale: a 16GB or 32GB Mac is not a realistic target for the currently advertised full-model local variants. [S22, S37] The unified-memory guide explains the memory-budgeting problem in more general terms.

Early self-hosting rough edges

Day-zero support is moving quickly. An SGLang issue documents a loop involving thinking mode plus the qwen3_coder tool parser. Another report describes QSA/SSM-path problems on an RTX PRO 6000 Blackwell SM120 configuration. [S27, S28]

llama.cpp support is also evolving, and an August 28 issue reports an N-gram-cache bug affecting speculative-decoding acceptance across requests in a development setup. [S29, S30]

These are not evidence that Alibaba’s hosted Qwen3.8-Flash endpoint is broken. They are evidence that a brand-new local architecture carries integration risk for teams self-hosting it immediately.

The license deserves a real read

The Flash-Next weights use Qwen Community License 1.0. It broadly grants rights to use, copy, modify, distribute, sell, deploy, host, fine-tune, and create derivatives, subject to conditions. The license adds attribution requirements at very large commercial scale and requires a separate license for certain Model-as-a-Service and AI Work Assistant commercial uses. [S36]

For commercial deployment, check the license text against your exact product. Calling the release simply “Apache-style open source” would be inaccurate.

Verdict

Qwen3.8-Flash is compelling when the workload combines long context, multimodal inputs, tool use, and low API pricing. It is not automatically the cheapest choice for short prompts; Qwen3.7-Flash can still cost materially less in the lowest context tier.

Flash-Next is a separate proposition: it is the more interesting release for architecture research, self-hosting, quantization, and Qwen4 experimentation, but it requires much more memory and a local serving ecosystem that is still stabilizing days after release.

The practical decision is therefore simple: choose hosted Qwen3.8-Flash for production API economics and long context; evaluate Flash-Next when control over weights and local inference matters enough to justify the hardware and integration cost.

Methodology

Research date: August 28, 2026. Thirty-seven sources were opened and reviewed, with first-party Alibaba/Qwen documentation prioritized for identity, limits, pricing, compatibility, and licensing; independent benchmark and implementation sources were used to test or qualify vendor claims. Pricing and deployment status are time-sensitive.