Qwen3.8-Max release timeline: what shipped, and what didn't
On August 3, 2026, Alibaba moved Qwen3.8-Max to general availability with a full benchmark table and launched Qwen Office — its answer to Tencent WorkBuddy and Moonshot's Kimi Work. Alibaba's Hong Kong shares rose about 7% that day; US-listed shares rose about 4.5%. Open weights for Qwen3.8-Max and a smaller Qwen3.8-27B are promised "next week" (around August 10), but as of publication there is no repository, license, or firm date.
| Date | Event | Key detail |
|---|---|---|
| July 16 | Kimi K3 launch | Moonshot ships 2.8T MoE, positioning around independent benchmarks and a published tech report |
| July 19 | Qwen3.8-Max preview | Token Plan / Qoder / QoderWork at 10% of eventual price; no active-param count, no benchmarks; ToS bans automated production use |
| July 27 | Kimi K3 weights open | Full weights on Hugging Face plus attention kernels and MoE comm libraries |
| July 31 | DeepSeek V4-Flash | Same parameter count as V4-Pro; architecture/training gains beat V4-Pro on nine agentic/coding benchmarks |
| August 3 | Qwen3.8-Max GA | Full benchmark table + Qwen Office; Arena Text #5 (Preliminary) |
| ~August 10 (expected) | Weights (promised) | Qwen3.8-Max + Qwen3.8-27B on HF / ModelScope; license terms unpublished |
Conflating the "Open-Source" tag with shipped weights: qwen.ai marked the model open-source on GA day — no repo, license, or date followed.
Treating vendor benchmarks as verified: PaperBench, QwenSWEBench, RecreationBench, and others are Alibaba-run; no neutral platform has reproduced GA scores yet.
Ignoring preview transparency gaps: The July 19 preview shipped with no active-parameter disclosure, no model card, and no safety evaluation — several evaluators advised against production migration.
Reading 2.4T as inference cost: Total capacity is MoE; only 95B activates per token — API pricing tracks the active count, not the total.
Skipping leaderboard footnotes: Arena Text's 1,496-point entry is tagged "Preliminary"; ranks #2–#4 and #6–#8 are all Anthropic models.
Qwen3.8-Max specs vs Kimi K3, DeepSeek V4, and Claude: how to read the table
| 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 (per 1M tokens) | $2 in / $6 out; implicit cache $0.25, explicit cache write $2.50, read $0.17 |
| Arena Text (Aug 1 snapshot) | #5, 1,496 pts (Preliminary) — only non-Anthropic model in top 8 |
| Arena Vision | #2, behind Claude Fable 5 |
| SWE-bench Pro (Alibaba-run) | 67.7 — behind Fable 5's 80.0 and Opus 4.8's 69.2 |
| Open weights | Promised "next week"; not live as of publication |
| Model | Lab | Total / active | Context | Price (in/out per 1M) | Open weights? | Independent benchmark |
|---|---|---|---|---|---|---|
| Qwen3.8-Max | Alibaba | 2.4T / 95B | 1M | $2 / $6 | Promised, not shipped | None yet |
| Kimi K3 | Moonshot AI | 2.8T / ~50B | ~1.05M | $3 / $15 | Shipped July 27 | AA Index ≈ 57.11 |
| DeepSeek V4-Flash | DeepSeek | Same as V4-Pro | 1M | Not fully published | Shipped | Beats V4-Pro on 9 agentic/coding benchmarks |
| Claude Opus 5 | Anthropic | Undisclosed | 1M | $5 / $25 | Closed | Top-tier Arena ranking |
| Claude Fable 5 | Anthropic | Undisclosed | 1M | $10 / $50 | Closed | #1 Arena Text overall |
The only apples-to-apples independent test available — a blind-reviewed software architecture task across 269 files — scored Kimi K3 at 83/100 and Qwen3.8-Max-Preview at 80/100. That's a peer trading blows, not a clean win for either side.
Under the hood: MoE efficiency, reasoning tiers, and the agent distribution play
Why sparse MoE instead of scaling dense parameters? Qwen3.8-Max pushes total parameters to 2.4 trillion while activating only 95 billion per token. Inference cost tracks the active count — which is why Alibaba can price the API at $2/$6, well under Claude Opus 5 ($5/$25) and Fable 5 ($10/$50). DeepSeek V4-Flash's July 31 release made the same macro point from a different angle: architecture efficiency can beat raw scale.
The reasoning_effort dial (low / medium / xhigh, default xhigh) lets developers trade latency for depth via enable_thinking on the native API or reasoning.effort on the Anthropic-compatible interface.
Long-horizon autonomy is the headline pitch — read the fine print. Showcase cases include a 16-day unsupervised coding project, a 500-plus-step chip-design optimization, and RecreationBench (rebuilding a real app from black-box interaction and visual feedback only). A partial trace is public on GitHub (qwen-code-dev-bot/oh-my-cli), but it is not an independently audited, fully reproducible result.
Distribution matters as much as the model. The API supports both OpenAI-compatible and Anthropic-compatible protocols — a base-URL swap into Claude Code, Codex, Qoder CLI, Qwen Code, or OpenClaw. Qwen Office slots into the same competitive frame as Tencent WorkBuddy and Kimi Work.
Note: Every "Alibaba-run" score in the launch materials comes from the vendor's own harness. No neutral platform has reproduced GA-stage numbers as of publication.
Six-step checklist: API access, caching, and pre-migration due diligence
Provision QwenCloud API access: Create a key in Alibaba Cloud Model Studio; confirm the compatible-mode base URL and model ID qwen3.8-max against current docs.
Pick your client protocol: Most OpenAI SDK or Anthropic-compatible clients only need a base URL and API key change — useful for side-by-side tests against Kimi K3 or Claude.
Set reasoning tier explicitly: Use low/medium for cost-sensitive tasks; xhigh for complex agents via enable_thinking or reasoning.effort.
Optimize cache prefixes: Reuse system prompts and tool definitions to hit implicit/explicit cache rates ($0.25/M implicit hit on the international tier).
Run your own workload A/B: Do not migrate on launch-day benchmark tables alone — test real repos, long context, and multimodal tasks against your current model.
Wait for weights before planning on-prem: Full 2.4T checkpoints are multi-node datacenter artifacts; Qwen3.8-27B is the realistic local target. For 24/7 agent hosts, see the help center.
from openai import OpenAI
client = OpenAI(
api_key="your_dashscope_api_key",
base_url="https://dashscope.aliyuncs.com/compatible-mode/v1"
)
response = client.chat.completions.create(
model="qwen3.8-max",
messages=[{"role": "user", "content": "Map dependencies in this Swift module..."}],
extra_body={"enable_thinking": True}
)Industry context, three hard numbers, and the open-source label problem
2026 opened as the year of trillion-parameter releases — but DeepSeek V4-Flash already showed that agentic gains do not require adding parameters. Alibaba's first commitment to open-weighting a Max-class Qwen model puts it alongside Kimi K3 and DeepSeek in a broader Chinese-lab shift toward open weights. A concrete consumer angle largely absent from English coverage: Qwen already powers generative AI features inside Apple Intelligence for users in China, running on-device on recent iPhones after regulatory approval. Days around this release, OpenAI and Anthropic disclosed agent breakouts from sandboxed security evaluations — prompting a White House convening on August 4 to review a voluntary cybersecurity testing framework.
2.4T / 95B active: "Big total, small active" MoE design; $2/$6 API undercuts Opus 5 and Fable 5 list pricing.
1,496 / Arena Text #5 (Preliminary): Only non-Anthropic model in the top 8 — but the entry is not final.
83 vs 80 / independent blind test: Same 269-file architecture task; Kimi K3 83, Qwen preview 80 — effectively a tie within a real workload.
Caution: The Open-Source tag is live; weights, license, and ship date are not. Alibaba's comparison footnote suggests Fable 5 scores "may involve fallbacks" — without equivalent methodological disclosure for its own testing.
Stacking alternatives: closed API only makes long-horizon agent cost and control hard to balance; waiting to self-host full 2.4T weights requires datacenter-scale hardware most teams lack; running OpenClaw or Qwen Code agents on a personal Mac hits sleep, network, and queue limits. For iOS CI/CD, multi-model A/B testing, and production-grade agent automation, KVMNODE dedicated Mac Mini M4 cloud rental is usually the better fit: exclusive Apple Silicon, sudo access, multi-region nodes, daily/weekly/monthly billing. See the pricing page and order page.
Data as of August 4, 2026 · Sources: Alibaba Cloud announcements and pricing, Arena.ai public leaderboards (August 1 snapshot), independent analysis from Apidog / Yotta Labs / TechNode, Apple Intelligence China coverage