Qwen made Qwen3.8-Max available on August 3, 2026, as a hosted model on QwenCloud. In its release announcement, the team says downloadable weights will follow next week.
That timing matters. Developers can call the hosted model now, but they cannot yet inspect or self-host it from the promised weights. Qwen’s service specifications and performance results remain vendor-reported; the eventual artifacts and independent testing will allow broader verification.
What is available now
Qwen describes Qwen3.8-Max as a mixture-of-experts model with 2.4 trillion total parameters and 95 billion active parameters. The current model ID is qwen3.8-max.
The QwenCloud model page lists text, image and video input with text output. It advertises a 1-million-token context window, up to 991,000 input tokens in the standard configuration, up to 983,000 input tokens when thinking is enabled, and up to 131,000 output tokens.
QwenCloud currently lists prices of $2 per million input tokens and $6 per million output tokens, plus separate cache rates. Those figures are service prices, not a measure of the hardware cost to run the future open weights.
The release also exposes reasoning_effort settings of low, medium and xhigh, with xhigh as the documented default. Qwen says preserve_thinking is enabled by default.
QwenCloud documents OpenAI-compatible Chat Completions and Responses APIs as well as an Anthropic-compatible interface. Compatibility at the protocol level does not guarantee identical tool-calling, streaming, error or reasoning behavior, so existing integrations still need regression tests.
The open-weights boundary
Qwen says the Qwen3.8-Max weights will be published on Hugging Face and ModelScope next week. At publication time, that promised release is not yet available to inspect.
The distinction should remain explicit: Qwen has announced “open weights,” not a complete self-hosting package already in hand. Teams will need to inspect the eventual license, model card, checksums, supported serving stacks, available quantized variants and practical hardware requirements to determine how reusable the release is outside QwenCloud.
Developers interested in self-hosting should wait for those artifacts and verify them directly rather than treating the API launch as proof that the weights have shipped.
What Qwen reports about coding and long-horizon work
Qwen’s announcement emphasizes multi-day coding, research and tool-driven tasks. The company says one highlighted software project ran for more than 10 days; the linked oh-my-cli repository provides a public artifact associated with that demonstration.
That artifact is useful for inspection, but it is not an independent evaluation. Qwen produced the model and demonstration and reported the results. The release’s benchmark table also combines vendor-run evaluations, internal benchmarks and published scores produced under different harnesses, context limits and timeouts.
Those results can guide follow-up testing, but they do not establish that the model will reproduce the same reliability, cost or speed in another organization’s codebase or agent harness.
What developers should verify
Before changing a production model route, teams should test Qwen3.8-Max on representative tasks in their own harness and compare completion quality, tool use, latency, token use and failure recovery.
They should also confirm the service price, regional endpoint, context limits, cache behavior and built-in-tool availability for the account and region they intend to use. API compatibility should be tested at the level of streaming events, tool arguments, structured outputs, errors and reasoning controls—not only whether a basic request succeeds.
For self-hosting, the next checkpoint is the actual weight release. Teams should inspect the license, model card, hashes, supported runtimes and memory requirements once Qwen publishes them.
Qwen3.8-Max is available today as a hosted model with an advertised 1-million-token context window and a promise of downloadable weights. The API release and the future open-weight release are two different milestones, and the evidence for each should be evaluated separately.
