Qwen3.8-Max: point Claude Code or Codex at a 2.4T model
The 2.4T Qwen flagship posts frontier agentic scores at $2/$6 per Mtok with a 1M-token window, and open-weights both it and a 24GB-ready 27B this week.
Copy markdown
Two endpoints, zero rewrites
Qwen3.8-Max speaks both OpenAI's Chat Completions format and Anthropic's API protocol, so you point Claude Code, Codex, or your own SDK at it by swapping the base URL and model ID. A 2.4-trillion-parameter MoE (95B active) becomes a drop-in for code you've already shipped.
$2 in, $6 out — cache is 8x cheaper
API pricing is $2.00 per million input tokens and $6.00 out, with cached reads at $0.25 (implicit) or $0.17 (explicit) — roughly 8x off fresh input. For agent loops that re-read the same context every turn, that cache line is where your bill actually lands.
Frontier scores on the agent evals
It posts 86.6 on Terminal-Bench 2.1, 86.1 on OSWorld-Verified, 92.6 on GPQA Diamond, and 93.0 on PaperBench — squarely in the frontier tier for terminal and computer-use agents, not just chat.
A 1M window, 262K of it for thinking
Context runs to ~1M tokens: up to 991K input (983K with thinking on), 131K output, and a 262K reasoning budget. Enough to load a mid-size repo and still leave room for a long tool-calling chain.
Open weights — and a 27B for one 24GB card
Alibaba says both checkpoints go open-weight on Hugging Face and ModelScope this week: the 2.4T flagship and a compact Qwen3.8-27B that fits a single 24GB GPU. The 27B is the one most builders will actually self-host.
Before you build on it: the license isn't out
The open-weight license text hasn't been published yet, and reporting flags possible geographic limits echoing recent Chinese 'open' releases. If you're shipping a product on the weights, wait for the terms.