Highlights
- Production Qwen4 flagship: Alibaba released
qwen4-maxon 16 November with a 2.4T backbone, 71B activated parameters per token, native vision and 1M context - Lower cost per task: the official Work360 average cost fell from $4.43 to $1.67, a 62% reduction
- Provider pricing on APIYI: $1.50 input, $4.50 output and $0.075 cached input per 1M tokens, with eligible top-up bonuses
- Three endpoints, four groups: Chat Completions, Responses and Messages are open across Default, SVIP, CodexResponses and ClaudeCode
- Measured, not repeated: 325 pre-release calls plus 15 Default-group release checks; a 951,498-token request recalled markers from both the middle and end
Information verified on 2026-11-17 (UTC+8). Primary sources: Qwen’s announcement at
qwen.ai/blog?id=qwen4-max and the QwenCloud model page at qwencloud.com/models/qwen4-max. APIYI pricing comes from the public pricing API, with test evidence collected on the same day.Background
Qwen4-Max is the first flagship in the Qwen4 family. Its target is not just higher scores on isolated reasoning questions, but sustained professional work: reading large codebases, analyzing long documents, using tools, and iterating from execution feedback. Two themes define this generation: native multimodality and long-horizon execution. Text, image and video input share a 1M context window. Alibaba also introduced Live Delivery, where a long-running agent provides a usable version early and then expands it through user feedback rather than returning only at the end. APIYI listedqwen4-max on 17 November. Alongside provider-aligned pricing, we tested long context, tool calling, structured output, caching and all three endpoint formats.
Deep dive
Architecture: balancing capacity and inference cost
GDN + QSA
Gated DeltaNet compresses history into fixed-size states, while Qwen Sparse Attention selects relevant context at micro-block granularity to reduce long-sequence attention and indexing costs.
Gated Residual
Four parallel residual branches use dynamic read and write gates to improve information flow, model capacity and training stability.
N-gram Embedding
Context-based lookups add local token-pattern representations. The extra 328B parameters increase capacity with little additional inference computation.
Official efficiency results: savings beyond token rates
The last two rows matter most: cost depends on both the price per million tokens and the number of tokens required to finish a task. Qwen4-Max reduces both, producing Alibaba’s reported task-level savings.
Official specifications
APIYI testing: near-1M context and three endpoints
We ran 325 pre-release calls covering capabilities, parameters and error boundaries, followed by 15 release checks with a Default-group token.- Near-1M recall: 951,498 input tokens returned in 92.3 seconds, with both mid-document and tail markers recalled correctly
- Long-context coverage: 8K, 32K and 128K recall passed on all three endpoints; a 32K release check also passed
- Tool calling: single, parallel, named and streaming calls worked with full tool-result replay; Responses has a
tool_choice="required"exception noted below - Multimodality: image OCR, two-image color and shape recognition, and video descriptions were correct
- Structured output: strict JSON Schema passed on Chat and Responses, including parseable streamed JSON
- Stability: 12/12 Chat requests succeeded at 12-way concurrency, and long streams ended cleanly on all three endpoints
Choosing an endpoint and group
These are client-specific recommendations, not restrictions on other groups. Minimal requests on all three endpoints also passed with a Default-group token. Each token’s available groups are governed by its console configuration.
Getting started
Recommended scenarios
- Codebase analysis and coding agents: read project context at scale, then return tool results for further execution
- Legal, financial and medical document analysis: combine long documents and images while retaining evidence locations
- Structured extraction: constrain fields with JSON Schema rather than parsing free-form text
- Long-horizon research and content production: deliver a usable draft, then revise through iterative feedback
Quick start
reasoning_effort="none" for everyday chat or structured extraction to reduce reasoning tokens; select a higher tier for complex reasoning.
Four integration notes
- Forced tools on Responses: provider testing ignored
tool_choice="required"; use a named tool or Chat Completions instead. - Structured output on Messages: OpenAI-style
response_formathas no effect andmax_tokensis required. Use tool calling for strict structures. - Basic JSON mode:
json_objectguarantees parseable JSON, not a top-level object or a field structure. Usejson_schemafor strict requirements. - Cache accounting: API usage is not a billing statement. Verify hits and discounts in the individual console log’s billing detail.
Pricing and availability
Pricing
Per 1M tokens before top-up bonuses:
Prices match the international Qwen4-Max provider site and may change. Check Model pricing for current rates.
Top-up promotions
Top-up bonuses of 10%–20% provide an effective rate of roughly 83–91% of list price. The bonus tier depends on a single top-up amount and does not change the model’s listed unit price.Summary and recommendation
Qwen4-Max is a strong candidate for long-document, multimodal and tool-driven agent workloads. Its main advances are architectural efficiency, long-horizon execution and professional deliverables rather than a single leaderboard result. Alibaba’s reported 62% lower average task cost is a reminder to evaluate both token rates and task token consumption. Use Chat Completions for general integrations, CodexResponses for Codex, and ClaudeCode for Claude Code. Disable thinking for routine calls, raise it for complex reasoning, and use JSON Schema for strict structures. See the Qwen4-Max Overview for full parameters and test notes.Sources: Qwen’s official announcement at
qwen.ai/blog?id=qwen4-max and the QwenCloud model page at qwencloud.com/models/qwen4-max, retrieved 2026-11-17 (UTC+8). APIYI findings cover 325 pre-release calls and 15 Default-group release checks. Official benchmark figures are provider-published and were not independently rerun.