Qwen3.8 Flash is available on Novita AI as a multimodal serverless model and an early preview of the Qwen4 architecture. It is the more efficiency-oriented Qwen3.8 option: the model combines 125B total parameters with 6B activated parameters per token, accepts text, image, and video input, and currently lists $0.15 per 1M input tokens, $0.016 per 1M cache-read tokens, and $0.47 per 1M output tokens on Novita AI. If you need a long-context multimodal model without starting with the Qwen3.8 Max pricing profile, Flash is the variant to evaluate first.
Qwen3.8 Flash at a Glance
The following values come from the Novita AI model page checked on August 31, 2026. The page is the source of truth for availability and billing because model limits and token prices can change independently of an article.
| Field | Current value |
|---|---|
| Display name | Qwen3.8 Flash |
| Model ID | qwen/qwen3.8-flash |
| Provider | Alibaba / Qwen |
| Access | Novita AI serverless API |
| Endpoint families | chat/completions, anthropic, responses |
| Input modalities | Text, image, video |
| Output modality | Text |
| Context window | 1,000,000 tokens |
| UI display | 977K context |
| Maximum output | 131,072 tokens |
| UI display | 128K max output |
| Input price | $0.15 / 1M tokens |
| Cache-read price | $0.016 / 1M tokens |
| Output price | $0.47 / 1M tokens |
| Platform release | August 27, 2026 |
The distinction between the normalized limits and the abbreviated UI labels matters. Novita’s page currently shows 977K and 128K in the feature panel while exposing 1,000,000 and 131,072 as the corresponding model values. Use the live model page when implementing hard validation or estimating a maximum request size.
What the Qwen4 Preview Means
Qwen3.8 Flash is not presented as a generic small model. Its positioning is an early preview of the Qwen4 architecture, with a design that targets broad input coverage and efficient activation rather than simply maximizing dense parameter count.
The model has 125B total parameters but activates 6B parameters per token. That mixture-of-experts layout can be useful for workloads that need a wide model capacity while keeping per-token computation more focused. The architecture also includes a 51B N-gram embedding component and combines GDN with QSA hybrid attention. These are architecture facts, not a promise of a particular benchmark result or latency in your application.
The practical difference is the combination of three capabilities in one endpoint:
- Multimodal input: Send text, images, or video when the task depends on more than a text transcript.
- Long working context: Use up to 1M tokens for large documents, repository material, or long video-related context, subject to the request format and live platform limits.
- Switchable reasoning behavior: Thinking mode is enabled by default on the model page and can be turned off when a faster or more concise response is preferable.
This makes Flash distinct from a conventional budget chat model. It is better understood as an efficient multimodal preview for teams evaluating the next Qwen architecture family through a hosted API.
When Qwen3.8 Flash Is a Good Fit
Multimodal document and media analysis
Use Flash when a workflow needs to combine written instructions with visual evidence. Examples include reviewing screenshots alongside an incident report, extracting facts from image-heavy documents, or asking questions about a video without maintaining separate model integrations for each input type.
Long-context coding and agent workflows
The 1M-token context capacity gives developers room to test repository summaries, issue history, tool traces, and design documents in one working set. The model’s 6B activated-parameter figure also makes its efficiency orientation relevant when an agent needs broad capability across many turns but does not always need the largest sibling model.
Use the context window as a capacity ceiling, not a target. Start with the smallest context that preserves the information needed for the task, then measure answer quality, latency, and spend as the prompt grows.
Long-video understanding
Video input is a meaningful reason to evaluate Flash separately from text-first Qwen endpoints. A video analysis workflow can use the same model family for visual and textual questions, but production testing should use representative clips, frame rates, resolutions, and question types. The model page confirms video input support; it does not guarantee a fixed cost or latency for every video shape.
Cost-sensitive frontier-model evaluation
At the currently listed rates, Flash is substantially cheaper per token than the Qwen3.8 Max article’s listed launch pricing. That does not mean it is automatically the lower-cost choice for every request. Actual spend depends on input length, cache reuse, output length, retries, and the way an application represents images or video. Compare total task cost, not only the headline input rate.
When It Is Not the Best Default
Qwen3.8 Flash may be unnecessary for short classification, extraction, or simple chat prompts where a smaller text model meets the quality bar. Multimodal support is valuable only when the application sends multimodal evidence; otherwise, it can add model-selection and request-format complexity without improving the result.
It is also not a substitute for a benchmark in your own workload. The available model-page description establishes the model’s modalities, limits, architecture positioning, and access path. It does not establish that Flash will outperform another model on your codebase, language mix, video corpus, or tool-calling loop. Evaluate with representative prompts before choosing a production default.
Teams that specifically need the highest-capacity Qwen3.8 option should compare the requirements and pricing of Qwen3.8 Max on Novita AI. Teams that already know the model choice and need runnable request examples should use the Qwen3.8 Max API quick start as the integration pattern, then substitute the Flash model ID after validating its supported endpoint and message format.
Limits, Pricing, and Source of Truth
Novita’s model page currently lists three token prices for Qwen3.8 Flash:
- Input tokens: $0.15 per 1M tokens
- Cache-read input tokens: $0.016 per 1M tokens
- Output tokens: $0.47 per 1M tokens
The cache-read rate matters when an application repeatedly sends stable system instructions, long reference material, or other reusable context. It should not be treated as a guaranteed discount for every repeated prompt; confirm how your request is classified and billed in the current Novita account and pricing surfaces.
For limits, the same page exposes a 1,000,000-token context value and a 131,072-token maximum output value, while the visible feature panel abbreviates them to 977K and 128K. Keep request-level limits below the live values until your integration has handled errors and truncation behavior. Do not hard-code the article’s prices into a billing system.
How to Access Qwen3.8 Flash
The model is available through Novita’s serverless API. Use the exact model ID qwen/qwen3.8-flash; OpenAI-compatible clients use the Novita base URL https://api.novita.ai/openai. The model page also lists Anthropic and Responses endpoint families. The API key, authentication, request body, and response handling should follow the current Novita AI API documentation rather than being copied from a stale article.
For a first evaluation, keep the request text-only and short even though Flash supports image and video input. Then add one modality at a time, cap output tokens, and record token usage. This separates authentication or endpoint errors from multimodal payload issues and gives you a baseline for quality and cost.
Conclusion
Qwen3.8 Flash is a strong candidate for teams that want to evaluate an early Qwen4 architecture preview through a hosted multimodal API. Its distinguishing profile is the combination of text, image, and video input, a 1M-token context ceiling, switchable thinking behavior, and an efficiency-oriented MoE design with 6B activated parameters per token. On Novita AI, the current listed price is $0.15 per 1M input tokens, $0.016 per 1M cache-read tokens, and $0.47 per 1M output tokens.
Choose Flash when those capabilities match the workload. If the job is short and text-only, test a smaller endpoint; if the job requires the largest Qwen3.8 capacity, compare Max; if the goal is implementation syntax, use a quick-start guide. In every case, treat the live Novita model page as the source of truth before production rollout.
FAQ
What is the Qwen3.8 Flash model ID on Novita AI?
Use qwen/qwen3.8-flash.
Does Qwen3.8 Flash support images and video?
Yes. Novita’s model page lists text, image, and video as supported input modalities, with text output.
What are the current Qwen3.8 Flash prices?
As checked on August 31, 2026, Novita lists $0.15 per 1M input tokens, $0.016 per 1M cache-read tokens, and $0.47 per 1M output tokens. Recheck the live model page before production budgeting.
How large is the context window?
The current model value is 1,000,000 tokens. The feature panel abbreviates this as 977K, so integrations should use the live platform limits rather than assuming the display label is the raw value.
Is thinking mode enabled by default?
Yes. The model page describes thinking mode as enabled by default and switchable off when the application needs a more direct response.
Sources
- Novita AI Qwen3.8 Flash model page, checked August 31, 2026
- Novita AI Create chat completion documentation, checked August 31, 2026
