Dek: Alibaba's new flagship is cheaper per finished task than the model it replaces, and that's the number worth arguing about. Its benchmark sheet is more mixed than the launch post suggests.
What Alibaba put on the table with Qwen4-Max
Alibaba introduced Qwen4-Max on 16 November, served through Alibaba Cloud. Input costs $1.50 per million tokens and output costs $4.50. A cached token costs $0.075, the price that decides whether a repeated prompt is cheap.
The spec sheet is unusual in one respect. Total parameters come to 2.4 trillion, and 71 billion of them fire on any given token. The model takes images natively, without a separate vision adapter, and reads a 1M-token context window.
Alibaba isn't claiming the best score on every board. It's claiming the same work for less money, which is a different argument and a harder one to check.
The mixture-of-experts design behind 2.4 trillion parameters
A mixture-of-experts spreads the network across many specialist blocks and routes each token through a small subset. Total parameters measure what you have to store and serve. Active parameters measure what you pay to compute, so 71 billion of 2.4 trillion is the number that ends up on an invoice.
Roughly 3% of the model runs per token.
Same bill, less money.
The rest waits.
Three design choices sit behind the efficiency claim. A 328-billion-parameter N-gram embedding layer acts as a cache of common token sequences, so routine phrasing gets retrieved rather than recomputed from scratch. Residual connections run through four parallel gated channels instead of the single channel most models use. Attention is hybrid: three layers compress earlier context, and a fourth selects only the parts that matter.
Alibaba says training ran without loss spikes, a claim that matters more than it sounds because large mixture models tend to become unstable at this scale. The company also reports that Qwen4-Max-Base improved 5-13% over Qwen3.8-Max-Base in bits per byte while activating 25% fewer parameters per token.
What a 1M-token context costs in practice
Long prompts used to be expensive for a structural reason: attention compares every token with every other token, so the bill grows faster than the input does. Gated DeltaNet and QSA are Alibaba's answer, cutting hardware demand for long inputs and skipping data the prompt doesn't need.
A 1M-token context is long enough to hold a mid-sized codebase, a year of support tickets or a full contract set in one pass. Whether accuracy holds at token 900,000 is a separate question, and the launch post doesn't answer it. If you've been building retrieval to work around a smaller window, that plumbing gets shorter.
The price maths is easy to run. A full million-token input at $1.50 per million costs $1.50 for the pass, before the model writes a word. Cheap for one document, less cheap when an agent re-reads the same context forty times in an afternoon.
Work360 and Alibaba's 62% cost cut claim
Work360 is Alibaba's own benchmark, built around 721 specialised industrial fields, and it's the cleanest place to see what the company thinks it solved.
The same task that cost $4.43 on the previous model now costs $1.67. That's a 62% cut. The benchmark score rose too, from 47.74 to 52.69.
Token use fell by 35% on an internal knowledge-work benchmark, and by about 49% on Terminal-Bench 4.0, the test that measures whether an agent can finish command-line tasks on its own. Every figure there comes from Alibaba, and the benchmark belongs to the vendor being measured.
Where Qwen4-Max trails Kimi and Tencent
The external picture is less flattering on coding. On Terminal-Bench 4.0, Qwen4-Max lands at 48% while Kimi K3.1 Pro reaches 67%, a gap wide enough to change which model you'd hand a refactor job.
FrontierCode 2.0 puts it at 26%, level with Kimi and behind Tencent Hy4 at 31%. HLE-Diamond gained 21 points to reach 45%, which is real progress and still behind Opus 5.5 and the GPT models on Humanity's Last Exam, the test built from expert-written questions across disciplines. HealthBench Professional has it behind the GPT models and slightly ahead of Opus 5.5.
So the honest summary is a model that improved a lot on reasoning-heavy academic tests, stayed mid-pack on agentic coding, and won on price.
Qwen4-Max pricing per task against the field
Artificial Analysis scores it at 43 on its Intelligence Index, nine points above Qwen3.8-Max and three behind Kimi K3.1 Pro at 46. The cost column is where the ranking flips.
Model | Intelligence Index | Cost per task |
|---|---|---|
Qwen4-Max | 43 | $3.27 |
Kimi K3.1 Pro | 46 | $3.68 |
Qwen3.8-Max | 34 | $10.02 |
Tencent Hy4 | 35 | $1.03 |
How many tokens does your own agent burn per task? That number, more than the price list, decides whether Qwen4-Max is cheaper for you. Per-task cost is the right measure here, and per-token prices across models aren't comparable when one model spends three times as many tokens reaching the same answer. Three points of index behind Kimi for slightly less money is close to a wash. Tencent's Hy4 costs a third as much and gives up eight index points, which is a different trade for a different buyer.
Live Delivery and the open-sourced agent plugin
The feature Alibaba is pushing hardest has nothing to do with benchmarks. Live Delivery returns a usable first draft from a long agent run, then lets the person redirect the work while it continues instead of waiting for a final answer they'll reject.
The Qwen-Live-Delivery-Plugin is open source, so the loop can be rebuilt around other models. Alibaba cites SANF and Pacific Union cutting manual programming time for one workpiece to about 24 minutes, down from roughly 475. It also describes a finance deployment that calls 90 MCP tools and returns Excel and PowerPoint files with numbers that trace back to source rows.
Those are partner stories, not case studies with published methods. Treat them as direction rather than proof.
What Alibaba still hasn't shipped with Qwen4-Max
Two gaps matter for anyone planning around this model. Qwen4-Max itself is proprietary, served only through Alibaba Cloud, so there are no weights to run on your own hardware. Alibaba says smaller open-weight models built on the same Qwen4 technology are coming, and hasn't shipped them.
The second gap is verification. Cost-per-task claims depend on token counts, and token counts depend on the test rig, the prompt and the retry policy. A cheap token only helps if the agent finishes the job, and the finishing is exactly what the vendor's own benchmarks can't settle.






