Nous Research Offers Qwen4-Max Free for 5 Days With 1M Token Context
Alibaba's 2.4T-parameter flagship with native vision and a 1M-token context is free to try on Nous Portal for five days through an Inferact partnership.
Takeaways
- Nous Portal is hosting Qwen4-Max free for five days via an Inferact and Alibaba partnership.
- Model ID
qwen/qwen4-max:freeexposes the full 1M-token context window and native vision. - Architecture: 2.4T-parameter MoE backbone with 71B parameters active per token, roughly 25% fewer than Qwen3.8-Max.
- Standard QwenCloud pricing is $1.50 input and $4.50 output per million tokens, undercutting Western flagships.
- Alibaba claims parity with top Anthropic and OpenAI models on coding; independent benchmarks still show a gap.
- Qwen4-Max is closed-weight; only smaller Qwen4 variants are slated for open-source release.
Nous Research has added Alibaba’s Qwen4-Max to Nous Portal for a five-day free trial. The model is served through Inferact’s platform using vLLM, an inference engine for running large language models, as part of a promotion with the Qwen team.
The free listing uses the model ID qwen/qwen4-max:free and requires no usage payment during the promotional window. The announcement defines the duration as five days without providing start and end timestamps, so developers should confirm that the free listing remains active before beginning a long evaluation.
One million tokens, 71 billion active
Qwen4-Max is the first model in Alibaba’s Qwen4 generation and succeeds Qwen3.8-Max. According to the Qwen announcement, it combines a 2.4 trillion-parameter Mixture-of-Experts backbone with native support for text and images.
Its sparse architecture activates 71 billion parameters for each token rather than using the entire backbone for every request. That is about 25% fewer active parameters than the predecessor’s 95 billion, reducing inference work while Alibaba reports higher overall capability.
| Feature | Details |
|---|---|
| Context window | Up to 1,000,000 tokens |
| Architecture | 2.4 trillion-parameter Mixture-of-Experts model |
| Active parameters | 71 billion per token |
| Inputs | Text and images |
| Weights | Closed; unavailable for download, self-hosting or modification |
| Target workloads | Coding, professional analysis and long-running agents |
Alibaba also lists 3D modeling, game development, precision manufacturing, finance, law, healthcare and scientific research among its target domains. Those workloads depend on more than context length, so evaluations should measure tool use, retrieval accuracy, latency and performance across long conversations.
Free tier, paid fallback
The Portal exposes separate free and metered listings, allowing developers to test the promotional model without assuming that requests will later move to a paid tier automatically. Teams planning continued use should select the paid model deliberately and apply spending limits.
| Provider | Input per 1M tokens | Output per 1M tokens | Availability |
|---|---|---|---|
| Nous Portal promotional tier | $0 | $0 | Five-day window |
| Nous Portal metered tier | $1.43 | $4.28 | Paid listing |
| QwenCloud | $1.50 | $4.50 | Official published price |
Connect, then verify metadata
Developers can use an OpenAI-compatible client or an agent framework such as Hermes Agent. A basic setup should follow these steps:
- Sign in to Nous Portal and confirm that the free listing is active.
- Create an API credential and copy the endpoint shown by the Portal.
- Configure the client with
qwen/qwen4-max:freeas the model ID. - Check the resolved context limit before sending a long request.
- Record any rate limits, timeouts and tool-call restrictions during testing.
A reported Hermes Agent issue can reduce the effective context window when a Portal metadata probe fails temporarily. Under that condition, a model with the :free suffix may inherit the generic Qwen-family limit of 131,072 tokens instead of Qwen4-Max’s advertised 1,000,000-token limit.
That fallback can trigger premature context compaction, where the agent summarizes or removes earlier messages to stay within its assumed limit. Long-horizon evaluations should inspect the resolved model metadata and compaction logs before treating a run as evidence of full-context performance.
Why the outside endpoint helps
The Portal gives developers another managed route to Qwen4-Max without configuring Alibaba’s cloud service directly. Because Alibaba has kept the flagship model’s weights closed and plans open releases only for smaller Qwen4 variants, third-party API access is the only practical option for teams that cannot use QwenCloud.
Serving the model through a familiar API also makes comparative testing easier. Existing harnesses can send the same prompts, tools and evaluation cases to Qwen4-Max, Claude, GPT or Gemini while holding generation settings and token budgets constant.
Benchmark claims need scrutiny
Alibaba presents Qwen4-Max as competitive with leading systems from Anthropic and OpenAI. The company says it placed near the top of coding evaluations and surpassed Claude Fable 5.1 and Moonshot’s Kimi K3.1 Pro on selected tests.
Independent benchmark coverage remains more measured and still reports gaps against some leading closed models. Alibaba also says Qwen4-Max reduced average task costs by 62% compared with Qwen3.8-Max in its Work360 evaluation, though results from a company-designed benchmark may not transfer to other workloads.
Reproducible comparisons require the benchmark version, prompts, sampling settings, tool configuration, scoring method and total token use. Cost should be calculated from successful task completion rather than token prices alone, since retries and longer outputs can erase a lower per-token rate.
Use the window for controlled tests
A useful five-day evaluation should run the same cases already used for other models and capture more than answer quality. Teams should track:
- Task success and coding test pass rates
- Tool-call accuracy and recovery from failed calls
- Long-context retrieval across different document positions
- Image understanding on relevant production inputs
- Latency, timeouts and request failures
- Input, output and hidden retry costs at paid rates
- Context compaction and memory behavior in agent runs
Teams that require downloadable weights, local inference or model modification cannot use Qwen4-Max for those purposes. Teams comparing managed APIs can use the promotion to test its million-token context, vision support and agent performance before deciding whether the metered Portal or QwenCloud pricing fits their workload.