Novita AI LLM, Qwen

Qwen4-Max on Novita AI: Native Vision, 1M Context, and Launch Pricing

Qwen4-Max on Novita AI: Native Vision, 1M Context, and Launch Pricing

Qwen4-Max is available on Novita AI as a serverless API under the model ID qwen/qwen4-max, priced at $1.413 per 1M input tokens and $4.24 per 1M output tokens. Novita lists a 977K-token context window, a 393,000-token maximum output, text, image, and video input, and text output through OpenAI-compatible and Anthropic-compatible endpoints. It is the successor flagship to Qwen3.8-Max, not a checkpoint update of it: a new generation in the Qwen4 family with its own listing, price, and limits. This page covers what is live on Novita AI, how it differs from the previous flagship, and who should evaluate it first.

Qwen4-Max Specs and Pricing on Novita AI

FieldDetails
Display nameQwen4-Max
Model IDqwen/qwen4-max
Base URLhttps://api.novita.ai/openai and https://api.novita.ai/anthropic
Endpoint familychat/completions, responses, Anthropic-compatible
Input modalitiesText, image, video
Output modalityText
Context window977K tokens displayed; 1,000,000 nominal
Max output393,000 tokens; the model page displays 384K
Input price$1.413 per 1M tokens
Cached input read$0.071 per 1M tokens
Cache write$1.767 per 1M tokens (5 minutes), $2.827 per 1M tokens (1 hour)
Output price$4.24 per 1M tokens
Release statusServerless; listed November 16, 2026
Best fitLong-horizon agents, multimodal document and video analysis, long-context coding, large deliverables

Specs and pricing from the Novita AI model page, checked November 16, 2026. The page displays 977K context and 384K output; the nominal context limit is 1,000,000 tokens and the API sample uses 393,000 max_tokens.

Key Takeaways

  • Qwen4-Max is the new Qwen4-family flagship, not a rename of Qwen3.8-Max. The two are separate Novita listings with different model IDs, prices, and limits.
  • Vision is native to the hosted endpoint. Text, image, and video input all resolve to text output through the same request.
  • The output ceiling is the clearest upgrade. Novita lists 393,000 max output tokens, against 128,000 on the Qwen3.8-Max listing.
  • Per-token pricing is lower than the previous flagship. $1.413 input and $4.24 output per 1M tokens, with cache reads at $0.071.
  • The model is a 2.4T-parameter MoE with 71B activated parameters. Novita lists reasoning, structured outputs, function calling, and serverless delivery.

What Is Qwen4-Max?

Qwen4-Max is Qwen’s new-generation flagship for long-horizon, production-grade work. Qwen describes a 2.4T-parameter MoE model with 71B activated parameters, a million-token context, and native visual understanding that runs through planning, execution, and verification. Novita’s model page adds that it is positioned to drive long-cycle tasks and deliver complete outputs, with a Live-Delivery mechanism for self-planning and closed-loop iteration.

The practical version is narrower. This is a model for work that needs many connected steps rather than one turn: a coding effort against a large repository, a workflow that checks its own output against a specification, or an analysis of hundreds of document pages and hours of video without dropping context.

Two details matter more than the parameter count. Vision is native, so images and video go into the same request as the instruction. And 393,000 output tokens fits a full report or a large code change in one response, removing the stitching smaller output caps force on pipeline design.

Qwen4-Max vs Qwen3.8-Max: A Successor Flagship, Not an Update

Qwen4-Max and Qwen3.8-Max are different models in different generations. Qwen3.8-Max remains its own listing, so a request to qwen/qwen4-max does not fall through to qwen/qwen3.8-max. If your configuration still names the older ID, nothing changes until you switch it and replay your evaluation.

The differences that affect billing and request design:

  • Output ceiling. Qwen4-Max lists 393,000 max output tokens; Qwen3.8-Max lists 128,000, with 131,072 in its API sample.
  • Price. $1.413 input and $4.24 output per 1M tokens, against $2 and $6 for Qwen3.8-Max. Cache reads are $0.071 against $0.25.
  • Generation. Qwen4-Max is the new-generation MoE flagship of the Qwen4 family; its architecture and long-horizon behavior are new, not an incremental checkpoint.
  • Vision. Novita’s current Qwen3.8-Max listing also shows text, image, and video input, so treat vision as a shared capability. The reason to move is generation quality, output size, and cost per token.

The Qwen3.8-Max launch guide documents the older profile for a side-by-side test.

Qwen4-Max API Access on Novita AI

Qwen4-Max runs serverless, so there is no deployment step: create an API key, point a client at the base URL, and pass the model ID. Novita lists two base URLs for the same model, https://api.novita.ai/openai for OpenAI-compatible clients and https://api.novita.ai/anthropic for Anthropic-compatible ones, with chat/completions and responses endpoint families. The model page sample uses the chat completions shape with max_tokens: 393000.

The listed features are reasoning, structured outputs, function calling, and serverless delivery, which covers most agent loops without extra infrastructure. For request fields and content-part syntax, use the Create Chat Completion reference. For a worked request example, the Qwen3.8-Max API quick start uses the same chat completions shape with a different model ID.

When to Use Qwen4-Max

Qwen4-Max fits workloads where the task, not a single turn, is the unit of work.

  • Long-horizon agents. Work that spans many tool calls, checkpoints, and revisions, where the deliverable is a finished artifact rather than a first draft.
  • Document and video analysis. Contracts, papers, dashboards, screen recordings, and long-form video can go in as image or video content and be analyzed in one request.
  • Large single-shot outputs. A 393,000-token ceiling fits a full technical report or a broad refactor without chaining sub-calls.
  • Repository-scale coding. A 977K context holds source trees, issue history, and test output together, reducing aggressive context trimming.
  • Flagship work on a lower per-token budget. At $1.413/$4.24 with $0.071 cache reads, the pricing profile undercuts Qwen3.8-Max, so teams already paying flagship rates can test it without a cost-model change.

When Not to Use Qwen4-Max

Qwen4-Max is the wrong default in these cases.

  • Short, high-volume requests. Classification, routing, and extraction from short text rarely need a 2.4T-parameter model; a smaller Qwen endpoint costs a fraction per call.
  • Hard interactive latency budgets. A flagship MoE with 71B activated parameters is not the low-latency option; measure time to first token before routing a UI through it.
  • Text-only work with no context pressure. If a smaller model passes your tests, the larger output ceiling and vision path add cost without value.
  • Production decisions without a track record. The listing is new, so there is no long production history to borrow. Run your own evaluation set first.
  • Pipelines that need low variance. When repeatability matters more than peak quality, a brand-new flagship is the wrong default.

How Qwen4-Max Fits Your API Workflow

Moving to Qwen4-Max is a configuration change plus an evaluation gate. Keep the model ID in configuration rather than scattered through code, move one environment to qwen/qwen4-max, and replay a fixed task set covering your hardest long-context, agentic, and multimodal cases. Compare task success, tool-call validity, retries, latency, and total token cost against Qwen3.8-Max or whatever you run today.

Two cost traps are worth measuring early. Image and video inputs consume tokens in ways text does not, so a vision prompt that looks cheap per request can move a budget at volume. Cache reads at $0.071 per 1M tokens are the lever for repeated long prompts; if an agent resends the same repository summary every turn, confirm cache hits instead of assuming them. Keep the previous model ID as a rollback path until the new endpoint clears your acceptance criteria.

Conclusion

Qwen4-Max is the first test worth running when your workload is long-horizon, multimodal, or limited by the previous flagship’s 128K output ceiling. It costs less per token than Qwen3.8-Max and takes images and video in the same request as text. It is not the default for short prompts or latency-critical interfaces, and it is too new to adopt on reputation alone. Start with a fixed evaluation set, watch multimodal token cost and cache behavior, and keep a rollback path to qwen/qwen3.8-max until the numbers justify the switch.

FAQ

What is the model ID for Qwen4-Max on Novita AI?

The exact model ID is qwen/qwen4-max. Use it in the model field with the OpenAI-compatible base URL https://api.novita.ai/openai, and confirm the current listing on the Qwen4-Max model page.

How much does Qwen4-Max cost on Novita AI?

As of November 16, 2026, Novita lists $1.413 per 1M input tokens, $4.24 per 1M output tokens, $0.071 per 1M cached input reads, and cache writes at $1.767 (five minutes) or $2.827 (one hour) per 1M tokens. Check the Novita AI pricing page before a production cost commitment.

Is Qwen4-Max the same as Qwen3.8-Max?

No. Qwen4-Max is a different model: Qwen3.8-Max is the previous flagship and remains a separate listing. It has its own ID, price, and limits, and requests to one do not route to the other, so update the model ID explicitly when you switch.

What are the context window and maximum output?

Novita displays a 977K-token context window and a 384K maximum output; the nominal context limit is 1,000,000 tokens and the API sample uses a 393,000 max_tokens value. Plan budgets against the lower displayed figures when requests sit near the limit.

Does Qwen4-Max accept image and video input?

Yes. The Novita listing shows text, image, and video input with text output, and the model page describes native visual understanding across planning, execution, and verification. Send one representative image or video segment through the standard content-part format before relying on it in production.

Is Qwen4-Max a good fit for coding agents?

It is a reasonable candidate for repository-scale, long-horizon coding: the 977K context holds source, issues, and test output together, and the 393,000-token output ceiling allows large single responses. Test it on your own failing tasks, because a new flagship stays an evaluation candidate until your acceptance criteria pass.

Can I call Qwen4-Max through the Anthropic-compatible endpoint?

Yes. Qwen4-Max is served through two base URLs: https://api.novita.ai/openai for OpenAI-compatible clients and https://api.novita.ai/anthropic for Anthropic-compatible ones. The model ID stays qwen/qwen4-max, and the model page shows the chat/completions and responses endpoint families, so the same model can be addressed without switching transports. Confirm the current endpoint list on the model page before wiring a client.

How is cached input billed for Qwen4-Max?

Novita prices cached input reads for Qwen4-Max at $0.071 per 1M tokens, against $1.413 per 1M uncached input tokens, and lists cache writes at $1.767 per 1M tokens for five minutes or $2.827 per 1M tokens for one hour. Long prompts that an agent resends across calls are where the lower read rate matters; confirm current rates on the Novita AI pricing page before budgeting.

Related Posts