Aug 3 GA · 2.4T / 95B active · Arena Text #5 · weights promised next week · vendor-run benchmarks
Short answer: not yet. On August 3, 2026, Alibaba made Qwen3.8-Max generally available through its cloud API and tagged it "Open-Source" on qwen.ai. As of this writing, there is no Hugging Face or ModelScope repository, no license terms, and no firm date—only a promise that weights for Qwen3.8-Max and Qwen3.8-27B will land "next week." This article maps the timeline, published specs, cross-vendor comparison, open-source label controversy, five-step adoption path, and FAQ—plus how to validate Agent workflows on a remote Mac when OAuth and macOS permission dialogs matter.
Open-Source tag shipped before weights: qwen.ai labels the model open, but no repo or license exists yet
Every benchmark is vendor-run: PaperBench, QwenSWEBench, RecreationBench, and more use Alibaba's own harness; Arena ranks it "Preliminary"
Active-parameter disclosure lagged: the July preview disclosed nothing; GA finally revealed 95B active
Full local deployment is unrealistic: a 2.4T checkpoint is a multi-node datacenter artifact even with MoE sparsity
Agent validation needs a GUI: OpenClaw, Claude Code, and Qoder CLI often hit macOS permission prompts that pure SSH cannot resolve
| Date | Event |
|---|---|
| July 16 | Moonshot AI releases Kimi K3 (2.8T MoE) with independent benchmarks and a technical report |
| July 19 | Qwen3.8-Max preview at 10% of eventual pricing; no active-param count, no benchmark table, ToS bans automated production use |
| July 27 | Kimi K3 open weights ship on Hugging Face on schedule |
| July 31 | DeepSeek V4-Flash beats V4-Pro on nine agentic/coding benchmarks without adding parameters |
| Aug 3 | Qwen3.8-Max GA with full benchmark table; "Qwen Office" agent product launches; Alibaba shares +7% HK / +4.5% US |
| ~Aug 10 (expected) | Open weights for Qwen3.8-Max and Qwen3.8-27B promised; not live as of publication |
| Spec | Qwen3.8-Max |
|---|---|
| GA date | August 3, 2026 |
| Total / active parameters | 2.4T / 95B |
| Architecture | Sparse MoE + hybrid attention on Qwen3.5 base |
| Context window | 1M tokens (≈983K with thinking; 131K max output) |
| Modalities | Text, image, video |
| API pricing | $2 / $6 per million input/output tokens |
| Arena Text (Aug 1) | #5, 1,496 pts (Preliminary)—only non-Anthropic in top 8 |
| Arena Vision | #2, behind Claude Fable 5 |
| PaperBench (Alibaba-run) | 93.0 (+28.2 vs prior gen) |
| SWE-bench Pro (Alibaba-run) | 67.7—behind Fable 5's 80.0 |
| Open weights | Promised next week; not live |
Quotable: 2.4T/95B · $2/$6 per M · Arena Text #5 (Preliminary) · OpenAI + Anthropic compatible API.
Sparse MoE pushes total parameters to 2.4 trillion while activating only 95 billion per token. Inference cost tracks the active count, which is why API pricing undercuts Claude Opus 5 ($5/$25) and Fable 5 ($10/$50).
reasoning_effort (low / medium / xhigh) is exposed via enable_thinking or Anthropic-compatible reasoning.effort—a standard cost dial for agent models.
Long-horizon autonomy is the headline: a 16-day unsupervised coding project, a 500+ step chip-design task, and RecreationBench (rebuild an app from black-box interaction only). Partial traces exist on GitHub (qwen-code-dev-bot/oh-my-cli), but benchmarks are in-house.
Distribution play: plugs into "Qwen Office" and drops into Claude Code, Codex, Qoder CLI, Qwen Code, and OpenClaw with a base-URL swap.
| Model | Total / active | Context | Price (in/out per M) | Open weights? | Independent bench |
|---|---|---|---|---|---|
| Qwen3.8-Max | 2.4T / 95B | 1M | $2 / $6 | Promised | None yet |
| Kimi K3 | 2.8T / ~50B | ~1.05M | $3 / $15 | Shipped Jul 27 | AA Index ≈57.11 |
| DeepSeek V4-Pro | 1.6T / 49B | 1M | Not fully published | Shipped | SWE-bench Verified 80.6% |
| DeepSeek V4-Flash | Same as V4-Pro | 1M | Not fully published | Shipped | Beats V4-Pro on 9 benches |
| Claude Fable 5 | Undisclosed | 1M | $10 / $50 | Closed | Arena Text #1 |
The only apples-to-apples independent test (269-file architecture task, blind-reviewed): Kimi K3 scored 83/100, Qwen3.8-Max preview 80/100—a peer trading blows, not a clean win.
The Open-Source tag went live before any weights did—a marketing decision until a repo exists
Every benchmark is vendor-run; no neutral platform has reproduced GA numbers
A footnote implies Fable 5 scores "may involve fallbacks" without equivalent methodological disclosure
The July preview banned automated production use and shipped without a model card or safety evaluation
Note: This does not mean the model is weak—the one independent blind test suggests frontier-class performance. Treat launch-day leaderboard claims as vendor claims until weights ship and third parties catch up.
Enable the API via Alibaba Cloud Model Studio; validate OpenAI- and Anthropic-compatible endpoints and reasoning.effort tiers
Map your workload to published scores—long-horizon agents vs weak HLE (43.6)
Point OpenClaw or Qoder at the new base URL; run probe tasks against Kimi K3 / DeepSeek for latency and cost
Watch for Qwen3.8-27B on Hugging Face if on-prem deployment is the goal
Finish OAuth, browser DevTools, and macOS permission checks in a real GUI macOS session (see VNCMac below)
No. The API is live, but weights are not on Hugging Face or ModelScope. The Open-Source label currently describes intent, not a shipped artifact.
One independent blind test: Kimi K3 83/100 vs Qwen3.8-Max 80/100. K3 has public weights and third-party scores; Qwen offers lower API pricing and broader multimodal support.
For the full checkpoint, yes. The API sidesteps that. Qwen3.8-27B is the realistic on-prem target when weights ship.
Treat them as vendor claims. Wait for third-party reproductions or A/B test on your own workload. Arena ranks the entry Preliminary.
Qwen powers Apple Intelligence generative features in China on recent iPhones—embedded infrastructure independent of direct API use.
Qwen3.8-Max lands Arena Text at #5 with aggressive API pricing and OpenClaw-compatible endpoints—but weights are not live and benchmarks are not independently verified. If your daily driver is Windows or Linux and you need a real macOS GUI for Qwen Office, OpenClaw routing, or parallel iOS builds, buying hardware is expensive and SSH alone cannot click OAuth or permission dialogs.
Renting a VNCMac remote Mac gives you an hourly VNC desktop to validate Qwen3.8-Max Agent workflows in isolation. See Mac mini M4 plans and our Kimi K3 open-weight breakdown for context.
Sources: Alibaba Cloud announcements, Arena.ai, TechNode, Apple Intelligence China coverage. Verify pricing and open-weight status before production migration.