Qwen3.8-Max goes open-weight: a 2.4T flagship and a 27B to run
The first frontier-scale Qwen to go open: 2.4T on the API today, a 27B to self-host, weights landing this week. Also: Google's Imagen 4 API dies Aug 17.
Copy markdown
The API is live now; the open weights land this week
Qwen3.8-Max is a 2.4T-parameter mixture-of-experts model (about 95B active per token) with a 1M-token context, already callable on QwenCloud at $2 in / $6 out per million tokens ($0.25 cached) over OpenAI- and Anthropic-compatible endpoints. It's the first Qwen Max-class model to go open-weight, with the download slated to land on Hugging Face and ModelScope this week.
A 27B you can actually run at home
Alongside the 2.4T flagship, Alibaba is releasing a dense Qwen3.8-27B for people without a data center. It's sized to fit a single 24GB consumer GPU at 4-bit (roughly 14-16GB VRAM). Build this weekend: a fully local coding agent on your own box, with no API bill and no data leaving the machine.
Tuned for coding and cowork
Qwen's own numbers put Max at 67.7 on SWE-bench Pro and 86.6 on Terminal-Bench 2.1, within reach of GPT-5.6 Sol (88.8), and a demo drove an autonomous coding run for 16 days across 265 commits. It reads as a genuine terminal/agent contender if you trust vendor benchmarks; independent runs will settle it once the weights are out.
Check the license and availability before you commit
Qwen's open models normally ship Apache 2.0, but the final terms for 3.8 are still being confirmed, and some reports flag regional restrictions. As of now the official Hugging Face repos are still rolling out; community FP8 and NVFP4A16 quant slots are staged and waiting on the base drop, so don't wire it into production until the real weights land.
Elsewhere: Google's Imagen 4 API shuts off Aug 17
If you call Google's image API, the imagen-4.0 endpoints stop working August 17. Migrate to gemini-3.1-flash-image, which swaps generate_images() for generate_content(), drops the batch number_of_images parameter, and runs $0.067 per image (or $0.034 via the Batch API).