Qwen4-Max brings native vision to long-running agents
By AgentRiot Editorial
Alibaba pairs a million-token multimodal model with staged agent deliveries and lower API rates. Independent tests show why workload choice still matters.

An agent that spends hours building the wrong thing needs more than a larger context window. It needs to hand over something useful before the budget is gone. That is the most practical idea in Alibaba's Qwen4-Max launch: combine native visual understanding and long-context reasoning with a delivery loop that lets users inspect intermediate work and change direction.
The Qwen team's announcement (opens in new tab), released November 16 in China and November 15 in US Eastern time, introduces a sparse model with a 2.4-trillion-parameter backbone and a one-million-token context window. It also releases a Live Delivery plugin that preserves successive versions of an agent's work. The model is available through QwenCloud, with consumer and coding-product access through QwenStudio, QwenWork and Qoder.
There is a credible efficiency story here, but not a clean sweep of the agent benchmarks. Independent operators show useful knowledge-work performance alongside weaker terminal results and uneven legal scores. The buying question is whether Qwen4-Max can produce an acceptable, inspectable deliverable more cheaply in your workflow, rather than whether it can stay busy longer.
A sparse architecture built for long context
According to Alibaba's architecture description (opens in new tab), Qwen4-Max activates 71 billion backbone parameters per token and adds 328 billion N-gram Embedding parameters beyond its 2.4-trillion-parameter backbone. The embedding capacity should not be confused with parameters activated for every generated token.
Its attention stack combines three Gated DeltaNet layers with each Qwen Sparse Attention layer. DeltaNet compresses history into fixed-size states; sparse attention uses a lightweight indexer to select relevant micro-blocks of context. Four parallel residual branches have dynamic read and write gates. N-gram lookups add representations of local token patterns at the second layer with relatively little extra computation.
These are mechanisms for controlling the computational cost of capacity and long sequences. They do not establish that the model reliably retrieves every fact across a million tokens. Alibaba's base-model comparison reports 5–13% lower bits per byte than Qwen3.8-Max-Base while activating 25% fewer parameters, but those are in-house prediction tests, not measurements of completed agent jobs.
The model-specific API guide (opens in new tab) supports text, image and video understanding with text output. Native vision can let an agent inspect a rendered game or chart and revise its code; it does not mean this model generates images or video itself.
The same guide gives exact limits: 991,808 input tokens and 131,072 output tokens without thinking; 983,616 input tokens and up to 262,144 tokens for thinking plus the final answer with thinking enabled. The shared context is one million tokens. Input and output maxima are not allowances to add together without regard to that window.
For a multi-turn agent, conversation handling also matters. The guide's CoT Summary feature can return a readable reasoning summary alongside encrypted full reasoning. The summary is for display; later turns need the encrypted history returned unchanged to restore that reasoning. Replacing the retained history with a short explanation would lose the distinction between showing users progress and preserving the model's working context.
Live Delivery puts an artifact between planning and more spending
The Live Delivery plugin's launch README (opens in new tab) describes three cooperating parts. A skill guides the agent toward a minimal usable first version and subsequent revisions. An MCP service manages plans, state and complete version snapshots through ld_begin, ld_publish, ld_status and ld_done. Lifecycle hooks feed user feedback back into the task and can block premature completion while work remains active and budget is available.
A local browser workbench exposes progress, previews and downloads, new requirements, budget adjustment and a stop control. For browser-created tasks, it interrupts and resumes the same Codex thread. External CLI tasks receive feedback at supported tool or turn events, so feedback is not necessarily applied at the instant a user types it. Stopping waits for the current tool to finish.
Consider a financial research website. A useful first release might show an imported dataset and one working chart. A reviewer can then reject an incorrect currency conversion before the agent adds backtesting and presentation features. That is an illustrative workflow, not a hands-on result. The important change is that review happens against files that already exist rather than a progress message promising they will exist later.
The implementation is narrower than the launch's general-purpose teammate language. Plugin 0.1.0 supports Codex CLI on Linux and macOS with Node.js 22 or later; its README names Codex 0.166.0 and 0.167.1 as tested versions. Other harnesses remain planned work. It reuses existing model and authentication configuration; it includes no model or API key.
Browser-created tasks default to workspace-write, but the pinned implementation sets their approval policy to never (opens in new tab). Do not assume installation preserving CLI permissions means browser tasks retain interactive command approvals. Explicit full access removes the sandbox too. The workbench binds to all network interfaces by default (opens in new tab); its entry link grants control. Use a trusted network or set QWEN_LD_BIND_HOST=127.0.0.1 for local access or SSH forwarding. Version snapshots improve reviewability, not correctness or permission safety.
Independent results favor some jobs, not every agent
Artificial Analysis (opens in new tab) gives Qwen4-Max an Intelligence Index score of 43, rounded from 42.89, at xhigh effort. Its v5.0.1 index combines nine evaluations across knowledge work, software engineering, science and document reasoning. A composite score is useful orientation, but the component results matter more when choosing a model for a particular job.
The operator's corresponding model records show this comparison. Percentages are scores, not a common success rate across different benchmarks; AA-Briefcase uses Elo, while AutomationBench-AA reports a partial-credit score.
| Model and evaluated effort | Intelligence Index | Terminal-Bench 4.0 | AutomationBench-AA | AA-Briefcase v1.1 Elo |
|---|---|---|---|---|
| Qwen4-Max, xhigh | 42.89 | 48.48% | 72.56% | 1686.07 |
| Kimi K3.1 Pro, max | 46.45 | 67.17% | 73.30% | 1673.56 |
| GPT-6.1 Sol, max | 54.61 | 56.06% | 64.87% | 1554.87 |
| GPT-6 Astra, max | 56.85 | 59.09% | 68.49% | 1568.82 |
| Claude Opus 5.5, max, default fallback | 59.11 | 59.60% | 69.54% | 1800.84 |
Qwen beats Sol on the knowledge-work Elo and SaaS-workflow score, yet trails it on the composite and terminal work. Kimi's terminal advantage is substantial, while Qwen's small AA-Briefcase Elo advantage over Kimi sits within overlapping published confidence intervals. Opus leads Qwen on AA-Briefcase and terminal work but loses on AutomationBench-AA. Different effort labels and Opus's fallback configuration remain part of those evaluated systems, not interchangeable settings.
AA's terminal evaluation uses mini-SWE-agent on 66 tasks with three repeats per task, not the Live Delivery plugin. Its AutomationBench-AA results (opens in new tab) also distinguish Qwen's 72.56% partial-credit score from 47.18% strict success. A high rubric score is not the proportion of jobs fully completed.
The launch's cited Agents' Last Exam v1.1 leaderboard (opens in new tab) adds a long-horizon result. Its entry named qwen-4, using Codex at xhigh, passes 57 of 122 tasks, or 46.72%, the highest strict-pass rate among the 30 configurations listed on November 16. Its mean rubric score, 74.39%, trails Gemini 4 Carbon's 77.41% and Kimi 3.1 Pro's 74.44%. These systems use different harnesses and efforts; the entry does not identify an immutable model deployment or Codex version. The strict-pass lead is useful evidence, not a controlled model-only victory.
Vals' November 15 evaluation (opens in new tab) supplies a second independent view. Qwen4-Max places 11th of 38 on Vals Index v2.2 at 63.20%, up from Qwen3.8-Max-0902's 59.05% and ahead of Kimi K3.1 Pro's 62.52%. Vals ran it at xhigh, temperature 1.0, top_p=0.95 and top_k=20.
Finance is comparatively strong: 66.68% on Finance Agent v2.1 and 61.99% on the Vals Finance Index, eighth of 41 on each. Legal performance is mixed. It scores 21.25%, seventh of 39, on HLAB-Vals-1.1, but 41.35%, tied for 32nd of 95, on Legal Research Bench. A good score on one legal workflow does not validate another. Its 48.49% Terminal-Bench result is consistent with AA's rounded figure, while Vibe Code Bench reaches 86.66%, 15th of 129.
That makes the launch's legal, manufacturing and scientific demonstrations examples of particular systems, not evidence of general professional reliability. Alibaba describes financial workbooks with source hyperlinks and presentations mapped to underlying spreadsheet data. Those are useful audit features to demand in a pilot. Its drug-discovery example, however, prioritizes a molecule using computational metrics for further wet-lab validation; it does not establish clinical effectiveness.
Lower token rates are only part of the cost story
Alibaba reports (opens in new tab) a Work360 score of 52.69, against 47.74 for Qwen3.8-Max. Work360 is its own benchmark, developed around a taxonomy of 26 top-level domains, 182 subdomains and 721 specialized fields. That breadth describes the framework, not proof that the model meets professional standards in every field.
The accompanying domain chart improves in 24 of 26 scored fields, but Administration falls from 50.90 to 49.60 and Scientific Research from 55.83 to 55.57. Its source labels those values as plotting-equivalent scores; they do not provide a disclosed reconstruction of the headline aggregate. The generation's overall gain should not be read as improvement in every domain.
The company says Work360 token usage falls approximately 24% at Xhigh effort and 35% at High. On Terminal-Bench 4.0, the reported reductions are about 48% and 49%. Combined with repricing, Alibaba puts average Work360 task cost at $1.67 rather than $4.43, a reported 62% reduction. That is a workload-specific vendor result, not a discount every customer receives.
AA's task-cost calculation (opens in new tab) is $3.27 for Qwen4-Max, versus $10.04 for Opus 5.5, $6.24 for GPT-6 Astra, $3.68 for Kimi K3.1 Pro and $1.27 for Sol at the settings above. AA weights evaluation costs and includes input, cache reads and writes, reasoning and answer tokens. Qwen's lower token tariff therefore does not make it cheaper than Sol on this workload.
Vals' index table (opens in new tab) makes the same warning concrete: Qwen averages $9.33 per test at 63.20%, compared with Sol's $3.61 at 63.94% and Kimi's $6.44 at 62.52%. Those figures belong to Vals' workload and accounting, not AA's. Cost per attempted test also differs from cost per accepted production deliverable.
API prices, cache writes and tool fees
The QwenCloud pricing guide (opens in new tab) lists these USD rates per million tokens. Thinking and non-thinking use the same rates, with no input-length pricing tiers for Qwen4-Max.
| Billing item | Singapore | Beijing / Global |
|---|---|---|
| Uncached input | $1.50 | $1.413 |
| Output, including thinking | $4.50 | $4.24 |
| Explicit cache creation, five minutes | $1.875 | $1.767 |
| Explicit cache creation, one hour | $3.00 | $2.827 |
| Explicit or implicit cache read | $0.075 | $0.071 |
For a synthetic Singapore workload spread across multiple requests and totaling one million uncached input tokens and 100,000 output tokens, including reasoning, the model charge is $1.95. If the same input volume is entirely valid cache reads, it is $0.525, before any cache-creation cost. These are token-arithmetic examples, not measured agent bills; they exclude tools, external services and retries.
Explicit caching (opens in new tab) charges more for creation than ordinary input. A 100,000-token prefix costs $0.1875 to create for five minutes and $0.0075 per subsequent full read in Singapore. Creating it once and reading it nine times costs $0.255, compared with $1.50 for ten uncached inputs of the same length. That assumes matching prefixes and a valid cache throughout; output is additional.
Prefixes must contain at least 1,024 tokens. Qwen4 explicit caching supports Chat Completions, Responses and Anthropic Messages, but not DashScope. Implicit hits are not guaranteed. Stable system prompts, tool definitions and reference documents are better reuse candidates than a constantly rewritten prefix.
The built-in-tool tariff (opens in new tab) adds Singapore web search at $10 per 1,000 calls and image search at $8. Beijing web search is $0.573411 and image search $6.881 per 1,000; Global regions list the Beijing web-search rate but no image-search offering. Web extraction and code interpreter are temporarily free, not permanently included. Function calling and MCP have no platform tool fee, although tool definitions consume input tokens and external providers can charge separately.
A free interpreter fee does not remove model-token charges: the tool can trigger multiple inferences, whose token use is aggregated. The interpreter's limitations (opens in new tab) also require streaming for Chat Completions and DashScope, and prohibit enabling interpreter and function calling together.
Access comes with endpoint and product boundaries
The current model guide (opens in new tab) identifies the production model as qwen4-max, with Chat Completions, Responses, Anthropic Messages and DashScope access. Thinking defaults to high with a 131,072-token budget; supported efforts are low, high and xhigh. Set effort explicitly in comparisons. max_completion_tokens covers reasoning and the final answer together, so budgeting only for the visible reply can truncate a result.
The guide supports structured output, parallel tool calls and preserved reasoning across tool turns; code interpreter requires thinking. It explicitly excludes batch inference, fine-tuning and model deployment.
Pay-as-you-go access covers Singapore/International, Beijing/China and Global workspaces in Frankfurt, Hong Kong, Tokyo and Virginia. Global inference can run in mainland China or other available regions (opens in new tab). Token Plan is purchased and accessed through Singapore but uses Global inference, including cross-border transfers, under its Individual terms (opens in new tab). A regional endpoint is not an inference-residency guarantee.
Token Plan Individual (opens in new tab) lists Lite at $6 for 20,000 monthly Credits, Essential at $10 for 45,000, Standard at $18 for 80,000 and Pro at $43 for 200,000. Credits vary with model, tokens, thinking and tools, so these are not fixed token bundles. Individual terms limit use to interactive programming and agent tools, excluding automation scripts and custom application backends. Token Plan (opens in new tab) is Singapore-only and uses separate keys and an endpoint from pay-as-you-go access.
The usage restrictions also apply to Team Edition (opens in new tab). Credits reset on the subscription anniversary (opens in new tab), not the calendar month, and unused quota does not roll over. Once plan and Credit Pack balances are exhausted, credit-consuming calls stop; there is no automatic pay-as-you-go fallback.
For sensitive documents, Token Plan's data policy (opens in new tab) explicitly excludes Team Edition conversation data from model training. Individual data use instead follows the service agreement, including section 5.2. Do not assume the Team guarantee covers an Individual subscription.
Max's weights are not publicly available, as AA's model listing (opens in new tab) records. Alibaba promises smaller open-source Qwen4 models later (opens in new tab); the open-source delivery plugin does not make Max open-weight.
For adoption, start with a bounded document-heavy task whose intermediate artifacts can be checked. Require citations that resolve, inspect spreadsheet formulas, and measure time and total spend to an accepted result at an explicit effort setting. Keep your existing terminal model in the comparison. Qwen4-Max's case is strongest when cheap reuse of a large context and early review reduce wasted work, not when a long run is mistaken for a successful one.

