Alibaba Locks Down Qwen4-Max Behind APIs at 62% Lower Cost
Alibaba's Qwen4-Max is a 2.4T-parameter agent model that cuts per-task costs by about 60%. It went live on You.com's Inference Router on day one, but its weights stay closed.
- Alibaba launched Qwen4-Max: 2.4T-parameter MoE, 71B active, 1M context, text/image/video input.
- You.com serves it from day one at $1.35/$4.05 per 1M tokens, below Alibaba's $1.50/$4.50.
- Alibaba claims Work360 cost per task fell 62% ($4.43 to $1.67) versus Qwen3.8-Max.
- It tops DeepSWE (75.7) but trails Kimi K3.1 and Opus 5.5 badly on Terminal-Bench 4.0 (48.5).
- Weights are closed, unlike Qwen3.8-Max. Smaller open Qwen4 models are promised but have no date.
- Training data was filtered to remove offensive-cyber material. Secure-code pass rate rose from 22% to 31.8%.
Alibaba has released Qwen4-Max, the first model in its Qwen4 generation. The 2.4-trillion-parameter flagship targets coding, office work and agents that run for hours. You.com added it to its Inference Router at launch.
The closed-weight release changes how developers can deploy Alibaba's largest model. Qwen3.8-Max was available for download, modification and self-hosting. Qwen4-Max is limited to Alibaba's products and third-party API providers, with fine-tuning and custom deployment unsupported.
Fewer experts fire per token
Qwen4-Max uses a sparse mixture-of-experts architecture. The model stores many specialist subnetworks, then activates a small selection for each token. Total parameters describe the model's capacity and storage footprint; active parameters more closely track the compute required by its expert layers.
| Feature | Specification |
|---|---|
| Total parameters | 2.4 trillion |
| Active parameters per token | 71 billion, down from 95 billion in Qwen3.8-Max |
| Context window | 1 million tokens |
| Supported inputs | Text, images, video and PDFs |
| Reasoning and output budget | Up to 262,144 tokens combined |
| API capabilities | Default reasoning, tool use and structured output |
How the model reduces compute
- Gated DeltaNet layers: This linear-attention design maintains compressed running memory, reducing the cost of processing long contexts.
- Qwen Sparse Attention: Each token attends to a selected subset of earlier tokens, shrinking attention work.
- N-gram embeddings: The model dedicates 328 billion parameters to lookup tables for common token sequences, trading additional storage for less computation through its main layers.
Alibaba says the pretrained base model activates 25% fewer parameters than its predecessor. Actual latency and throughput will also depend on attention costs, batching, quantization and the provider's serving stack.
Token savings drive the price case
You.com charges $1.35 per million input tokens and $4.05 per million output tokens, 10% below Alibaba's international list prices of $1.50 and $4.50. Both routes are cheaper than Qwen3.8-Max, which Alibaba lists at $2.00 for input and $6.00 for output.
| Model and provider | Input per 1M tokens | Cached input per 1M tokens | Output per 1M tokens |
|---|---|---|---|
| Qwen4-Max through You.com | $1.35 | $0.068 | $4.05 |
| Qwen4-Max through Alibaba | $1.50 | About $0.075 | $4.50 |
| Qwen3.8-Max through Alibaba | $2.00 | N/A | $6.00 |
Cached-input pricing applies when a provider can reuse an unchanged prompt prefix, such as a long system prompt or shared document. Cache behavior, minimum prefix lengths and retention periods can affect the realized discount.
Alibaba's Work360 test reports that average cost per task fell from $4.43 with Qwen3.8-Max to $1.67 with Qwen4-Max, a reduction of about 62%. The Work360 score rose from 47.74 to 52.69.
On Terminal-Bench 4.0, Alibaba reports about 151,000 tokens per attempt, down from 289,000, while the score increased to 48.5. Work360 showed smaller token reductions of about 24% at the highest reasoning setting and 35% at the next setting.
These efficiency figures come from Alibaba's test harness. Token use can change with prompts, stopping rules, tool implementations and retry policies, so production savings require measurement on the intended workload.
Code repair leads the benchmark sheet
Among 16 launch benchmarks, Alibaba reports four first-place results. The selected scores show strong code-repair performance alongside weaker terminal and software-reconstruction results.
| Benchmark | Qwen4-Max | Peer results |
|---|---|---|
| DeepSWE v1.1, agents fixing real code | 75.7 | Kimi K3.1: 75.5; GPT-6.1 Sol: 75.2; Claude Opus 5.5: 74.2 |
| Agents' Last Exam v1.1 | 46.7 | Claude Opus 5.5: 43.4; Qwen3.8-Max: 27.9 |
| Terminal-Bench 4.0 | 48.5 | Kimi K3.1: 67.2; Claude Opus 5.5: 59.6; GPT-6.1 Sol: 56.1 |
| ProgramBench, software reconstruction from observed behavior | 31.5 | Claude Opus 5.5: 65.0 |
Independent testing supports the Terminal-Bench result. Vals AI measured 48.48%, effectively matching Alibaba's figure, and ranked Qwen4-Max 15th among 65 models. The model placed 11th of 38 on the overall Vals Index, just ahead of Kimi K3.1 Pro.
Arena ranked Qwen4-Max as the leading Chinese model on its agent leaderboard. The model also debuted sixth globally on Code Arena's web-development ranking.
The Max line returns to API-only access
Alibaba released Qwen3.8-Max as the open-weight Qwen3.8-2.4T-A95B, its first downloadable model at Max scale. Qwen has also become the world's most-downloaded open-weight model family. Qwen4-Max ends that brief precedent for the flagship tier.
Developers can call Qwen4-Max through APIs or use it in Alibaba's applications. The weights cannot be downloaded, modified or run on private hardware. Alibaba's documentation also lists fine-tuning and custom deployment as unsupported.
Alibaba attributes the restriction partly to security. The company says it filtered training data to retain defensive material, such as vulnerability detection, while removing content intended to improve offensive capabilities. On its test of functional and secure code generation, the function-level pass rate increased from 22.0% to 31.8%.
API-only distribution also lets Alibaba enforce safeguards centrally and collect inference revenue. Customers inherit platform dependencies involving rate limits, regional availability, pricing and compliance rules, with no self-hosted fallback.
You.com provides the alternate endpoint
You.com serves Qwen4-Max through an OpenAI-compatible API. Its current upstream provider is Alibaba Singapore. You.com states that the route uses zero data retention and excludes customer data from model training.
An existing OpenAI client requires a You.com API key, the new base URL and the qwen4-max model identifier:
from openai import OpenAI
import os
client = OpenAI(
base_url="https://api.you.com/v1",
api_key=os.environ["YDC_API_KEY"],
)
response = client.chat.completions.create(
model="qwen4-max",
messages=[
{
"role": "user",
"content": "Refactor this module for testability.",
}
],
)
print(response.choices[0].message.content)The You.com playground provides browser-based testing and optional live web search for answers grounded in current sources.
Alibaba also offers the model through QwenCloud, Qwen Studio, QwenWork and the Qoder coding tool. Its open-source Qwen-Live-Delivery-Plugin allows long-running agents to deliver an initial version, continue improving it and accept user corrections before the run finishes.
Where Qwen4-Max fits
| Workload or requirement | Practical consideration |
|---|---|
| Hosted code-repair agents | DeepSWE performance and low output pricing support workload-specific evaluation. |
| Terminal-heavy agents | Peer models scored substantially higher, making direct comparison important. |
| Long software reconstruction | ProgramBench exposes a large gap from the leading result. |
| Self-hosting or fine-tuning | Qwen4-Max does not support either deployment path. |
| Long-context multimodal work | The one-million-token window accepts text, images, video and PDFs; accuracy and latency still require testing at production lengths. |
| Regulated data | Teams should review provider contracts, regional routing and retention guarantees before deployment. |
Preview retirement comes first
- Model ID migration: Alibaba says
qwen4-max-previewwill retire. Integrations should move toqwen4-max. - Smaller open models: Alibaba has named four Qwen4 tiers and promised open weights for smaller versions. Sizes, licenses and release dates remain unannounced.
- Broader independent tests: External evaluations will show whether the reported token reductions persist across different prompts, tools and stopping policies.
A production evaluation should track completed-task rate, median and p95 latency, tool-call failures, input and output tokens, retries and total cost per successful run. Published results align most closely with code-repair workloads. Terminal-heavy systems and teams requiring local weights have stronger reasons to compare alternatives or wait for the smaller open Qwen4 releases.