cd /news/large-language-models/alibaba-s-qwen4-max-slashes-agent-ta… · home › topics › large-language-models › article
[ARTICLE · art-170212] src=alphasignal.ai ↗ pub= topic=large-language-models verified=true sentiment=↑ positive

Alibaba's Qwen4-Max Slashes Agent Task Costs 62% With 1M-Token Context

Alibaba's Qwen team released Qwen4-Max, a proprietary mixture-of-experts flagship with native text, image and video input and a 1 million-token context window, priced at $1.50 per million input tokens and $4.50 per million output tokens. Alibaba reports Qwen4-Max uses up to 50% fewer tokens than Qwen3.8-Max, cutting average Work360 cost per task from $4.43 to $1.67, and scores 52.69 on Work360, a first-party benchmark spanning 721 professional fields that has yet to be independently reproduced. The model is live on the QwenCloud API, Qwen Studio, QwenWork and Qoder, and Alibaba has promised smaller open-weight Qwen4 models without a release date.

read6 min views1 publishedNov 16, 2026
Alibaba's Qwen4-Max Slashes Agent Task Costs 62% With 1M-Token Context
Image: Alphasignal (auto-discovered)

Alibaba's new flagship uses a fresh MoE design to cut token usage nearly in half while beating Qwen3.8-Max on productivity benchmarks.

  • Alibaba launched Qwen4-Max , a new MoE flagship with native vision and 1M-token context.
  • Pricing drops to $1.5/M input and $4.5/M output tokens, with implicit caching at $0.075/M.
  • Uses up to 50% fewer tokens than Qwen3.8-Max; average Work360 cost per task falls from $4.43 to $1.67.
  • Scores 52.69 on Work360, a new benchmark spanning 721 Tier-3 professional fields.
  • Live now on QwenCloud API , Qwen Studio, QwenWork, and Qoder.
  • Smaller open-weight Qwen4 models promised, with no release date yet.

Alibaba’s Qwen team has released Qwen4-Max, a new generation of its flagship proprietary model. According to the release post, it combines a redesigned mixture-of-experts architecture, native text, image, and video input, and a 1 million-token context window. Alibaba says the model matches or exceeds Qwen3.8-Max on its evaluations while using as much as 50% fewer tokens.

The full release follows a two-week preview that began November 11 through Alibaba’s Token Plan subscription, with reasoning enabled by default. The preview details preceded API access through QwenCloud and integrations with Qwen Studio, QwenWork, and the Qoder coding environment.

Alibaba puts cost per task first #

Agentic systems can invoke a model dozens of times while planning, calling tools, checking results, and retrying failed steps. Each iteration adds input, output, and reasoning tokens, so token efficiency can affect operating costs as much as the per-token rate.

Alibaba reports the following results from its internal evaluations:

Alibaba’s reported Qwen4-Max efficiency results
Evaluation Reported result
--- ---
Work360, Xhigh reasoning About 24% fewer tokens
Work360, High reasoning About 35% fewer tokens
Terminal-Bench 4.0 About 48% to 49% fewer tokens
Work360 average task cost $1.67, down from $4.43

Qwen4-Max also lowers the listed token rates from the earlier prices charged for Qwen3.8-Max:

Price per million tokens
Token type Qwen4-Max Qwen3.8-Max Reduction
--- --- --- ---
Input $1.50 $2.00 25%
Output $4.50 $6.00 25%
Implicit cache $0.075 $0.25 70%

Work360 needs outside validation #

Alibaba created Work360 to evaluate tool use and long-running professional tasks. Its hierarchy covers 26 top-level domains, 182 subdomains, and 721 specialized fields spanning industrial and professional workflows. Qwen4-Max scores 52.69, which Alibaba says is roughly 10% higher than Qwen3.8-Max.

The benchmark’s task-oriented design reflects how developers deploy models in agents, where success depends on planning, tool selection, and execution across multiple steps. Its first-party origin limits the strength of comparisons until independent evaluators reproduce the results. A circulating chart has assigned Qwen4-Max a score of 53 on the Artificial Analysis index, but Artificial Analysis did not list the model in its rankings at publication time.

Key architecture details remain undisclosed #

Alibaba has not disclosed Qwen4-Max’s total parameter count, active parameter count, or expert configuration. Its predecessor used a 2.4 trillion-parameter mixture-of-experts architecture with about 95 billion parameters active for each token. In this architecture, a router sends each token through a subset of specialized subnetworks, providing a large total capacity without activating every parameter during each step.

Reports from the preview described Qwen4-Max as a lean-activation model with a large n-gram memory layer, an auxiliary mechanism for handling recurring token sequences. Alibaba’s public description confirms a new mixture-of-experts design but does not provide enough detail to verify those architectural reports.

API compatibility eases migration #

Qwen4-Max accepts text, images, and video and supports tool calling, structured output, and web search. QwenCloud offers interfaces compatible with OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages. An existing client can begin migration by changing its API key, base URL, and model identifier, then validating tool schemas, streaming behavior, error handling, and token accounting.

A minimal OpenAI-compatible request looks like this:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_QWEN_KEY",
    base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",
)

response = client.chat.completions.create(
    model="qwen4-max",
    messages=[
        {"role": "user", "content": "Summarize this contract..."}
    ],
)

print(response.choices[0].message.content)

The model supports up to 393,216 output tokens, shared between reasoning and the final response. Production clients can control cost and latency by setting explicit output limits, selecting an appropriate reasoning-effort level, and terminating agent loops when progress stalls.

Tool-heavy workflows are the target #

Alibaba positions Qwen4-Max for workflows that require repeated tool use, long context, and multi-step execution. The announced use cases include:

  • 3D modeling and cinematic design, from blockout through rendering.
  • Game development.
  • Manufacturing workflows that convert sketches into production specifications.
  • Finance, legal, and healthcare knowledge work.
  • Scientific research, including autonomous design and synthesis of coolant materials.

Public preview reports described strong results on greenfield projects and less consistent performance when editing existing work or completing long sequences of dependent steps. Those reports are anecdotal, but they identify useful tests for teams evaluating the model against mature codebases and stateful agent workflows.

One side-by-side test reported that Qwen4-Max completed a coding task in about 2.5 minutes, compared with 17 minutes for Qwen3.8-Max. The same test reported 12 minutes for an animation task, compared with 19 minutes for the previous model. Task design, hardware, prompting, and reasoning settings can materially affect such latency comparisons.

Lower token use compounds the price cut #

Qwen4-Max’s economic case combines a 25% reduction in input and output rates with lower reported token consumption. That combination produced an approximately 62% decline in Alibaba’s reported average Work360 task cost, from $4.43 to $1.67.

The 70% cache-price reduction targets workloads that repeatedly send the same large prompt prefix, including retrieval-augmented generation, document review, and codebase agents. Implicit caching lets the provider reuse previously processed context at a discounted rate, making cache-hit rate a significant cost variable for long-context applications.

Open weights have no release date #

Qwen4-Max remains proprietary and available through hosted services. Alibaba says smaller Qwen4 models will receive open-weight releases, although it has not published a schedule. The announced family includes Qwen4-Max, Qwen4-Flash, Qwen4-Plus, and Qwen4-27B; parameter counts for the other variants remain undisclosed.

Qwen3.8’s open-weight models arrived several weeks after its cloud release, but that earlier cadence does not establish a timetable for Qwen4. Teams that require self-hosting will need to wait for model weights, licenses, hardware requirements, and quantization options before planning deployments.

Measure the full agent loop #

Teams comparing Qwen4-Max with Qwen3.8-Max or another frontier model can evaluate the operational claims by recording:

  • Task completion: Whether the agent reaches the required result without manual intervention.
  • Token use: Input, output, reasoning, and cached tokens for each completed task.
  • Tool reliability: Invalid calls, retries, timeouts, and recovery attempts.
  • Latency: Time to first token and total wall-clock time.
  • Cost: Total spend per successful task, including failed runs.
  • Quality: Domain-specific correctness, regressions, and edit accuracy.

Qwen4-Max is available through the QwenCloud API, Qwen Studio, QwenWork, and Qoder. Its main deployment claim is measurable: comparable or better task performance with fewer tokens and lower unit prices. Workload-specific tests will determine whether those savings survive real tool calls, retries, long contexts, and production constraints.

── more in #large-language-models 4 stories · sorted by recency
── more on @alibaba 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/alibaba-s-qwen4-max-…] indexed:0 read:6min 2026-11-16 · —