Open-Source label · 2.4T MoE · Arena #5 · Weights promised next week · Kimi K3 / DeepSeek comparison
If you are trying to answer whether Qwen3.8-Max is actually open source, or how it stacks up against Kimi K3 and DeepSeek V4-Flash, the short answer is: not yet on weights. On August 3, 2026, Alibaba GA'd a 2.4T-parameter flagship through its cloud API and tagged it "Open-Source" on qwen.ai — but no Hugging Face or ModelScope repo exists as of August 4. This article walks through the release timeline, vendor-run benchmarks, model comparison table, six-step API runbook, and the open-source label gap so you can make an evidence-based call. Data as of 2026-08-04.
Chinese frontier models shipped in rapid succession through July 2026. Understanding what actually landed — versus what was promised — is the only way to judge whether "frontier tier" claims hold up today.
July 16: Moonshot AI releases Kimi K3, a 2.8T-parameter MoE model built around independent benchmarks and a published technical report. See our Kimi K3 open-weight breakdown.
July 19: Qwen3.8-Max preview opens via Token Plan / Qoder / QoderWork at 10% of expected GA pricing — no active-parameter count, no benchmark table, and terms of service banning automated production use.
July 27: Kimi K3 ships open weights on schedule to Hugging Face, plus parts of its serving stack (attention kernels, MoE communication library).
July 31: DeepSeek releases V4-Flash, beating its own V4-Pro preview on nine agentic and coding benchmarks without increasing parameter count.
August 3: Qwen3.8-Max goes GA with a full benchmark table and companion agent product "Qwen Office." Alibaba's Hong Kong shares rose about 7%; US shares about 4.5%.
Expected around August 10: Open weights for Qwen3.8-Max and smaller Qwen3.8-27B promised on Hugging Face and ModelScope — no repository, license, or firm date as of publication.
Teams tracking this release often hit five blind spots:
Open-Source tag ahead of weights: qwen.ai marked Qwen3.8-Max "Open-Source" on GA day while Hugging Face and ModelScope still have no corresponding repository.
All benchmarks are vendor-run: PaperBench, RecreationBench, QwenSWEBench, and others come from Alibaba's own harness — no third-party reproduction of GA numbers yet.
Active params disclosed late: Alibaba revealed nothing about active parameters during the July preview, only disclosing "95B" at GA — a transparency gap independent evaluators flagged in mid-to-late July.
Local deploy fantasy: A 2.4T checkpoint needs multi-node datacenter hardware; individual developers should wait for Qwen3.8-27B.
Preview TOS blocked production: The July 19 preview shipped with terms banning automated production use, no model card, and no published safety evaluation — several evaluators advised against migrating production workloads on announcement alone.
| Spec | Qwen3.8-Max |
|---|---|
| GA date | August 3, 2026 |
| Total / active parameters | 2.4T / 95B |
| Architecture | Sparse MoE + hybrid attention, built on Qwen3.5 base |
| Context window | 1M tokens (≈983K with thinking enabled; 131K max output) |
| Input modalities | Text, image, video |
| API pricing | $2 / $6 per million input/output tokens (implicit cache $0.25, explicit cache write $2.50, read $0.17) |
| Arena Text Arena (Aug 1 snapshot) | #5 overall, 1,496 points (Preliminary) — only non-Anthropic model in top 8 |
| Arena Vision Arena | #2, behind Claude Fable 5 |
| PaperBench (Alibaba-run) | 93.0 (+28.2 vs prior generation) |
| 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 |
Every row marked "Alibaba-run" comes from the vendor's own launch materials. As of publication, no independent platform — Artificial Analysis, Arena.ai's official team, or otherwise — has published a reproduced score for the GA release.
Qwen3.8-Max API supports both OpenAI-compatible and Anthropic-compatible protocols, dropping into Claude Code, Codex, Qoder CLI, Qwen Code, and OpenClaw with a base-URL swap. Here is a six-step path from evaluation to production:
Define task scope and cost boundary: Separate API calls (billed on active parameters) from local privatization (wait for Qwen3.8-27B or datacenter-scale clusters). Long-context workloads should model cache hit rates against the bill.
Open QwenCloud / Model Studio account: Create an API key on Alibaba Cloud Model Studio and confirm rate limits and regional availability.
Pick interface protocol: OpenAI-compatible endpoint for existing SDKs; Anthropic-compatible endpoint exposes low/medium/xhigh reasoning tiers via the reasoning.effort field (default xhigh).
Configure reasoning_effort: Use low for simple tasks to control latency and cost; use xhigh for complex agent long-horizon work, or enable thinking mode via native enable_thinking.
Wire into agent host environment: Point base URL to Qwen endpoints and validate tool calling and multi-turn stability in OpenClaw / Claude Code — avoid running under preview terms in production.
Build A/B baseline and fallback routing: Compare Kimi K3 and DeepSeek V4 API latency, cost, and success rate on your own samples; configure primary/backup model routes and revisit defaults once third-party scores land.
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": "Analyze this code architecture..."}],
extra_body={"enable_thinking": True}
)
| Model | Lab | Total / active params | 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 (16 of 896 experts) | ~1.05M | $3 / $15 | Shipped July 27 | Artificial Analysis Intelligence 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 on Arena Text Arena overall |
Qwen3.8-Max pushes total parameters to 2.4T while activating only 95B per token — inference cost tracks the active count, not the total. That is why API pricing sits well under Claude Opus 5 ($5/$25) and Claude Fable 5 ($10/$50). Kimi K3 and DeepSeek disclosed active counts earlier; Alibaba waited until GA, which was a core part of the preview-stage transparency criticism.
In the only apples-to-apples independent test available — a third-party evaluator running Qwen3.8-Max-Preview and Kimi K3 against the same real-world software architecture task (269 files, blind-reviewed) — Kimi K3 scored 83/100 and Qwen3.8-Max scored 80/100. That is a peer trading blows, not one model dominating the other.
The Open-Source label went live before any weights did. qwen.ai marked Qwen3.8-Max "Open-Source" on GA day while the repository, license, and ship date remained unpublished. Every benchmark spans Alibaba's own harness; the comparison table footnote implies Fable 5 "results may involve fallbacks" without Alibaba publishing equivalent methodological detail for its own testing.
Qwen3.8-Max is genuinely frontier-class based on the one independent blind test available — but claims like "second only to Fable 5" currently rest entirely on Alibaba's own characterization until weights ship and independent leaderboards catch up.
API access covers Qwen3.8-Max inference, but iOS signing chains, Xcode local builds, Metal workloads, and 24/7 agent long-runs still need real macOS nodes — laptops cannot reliably host persistent OpenClaw / Claude Code agents, and VMs add performance tax plus EULA risk. For production iOS CI/CD and AI agent automation, VpsMesh Mac Mini cloud rental is usually the better fit: bare-metal Apple Silicon, root access, and predictable monthly cost pair cleanly with Qwen API in a cloud-inference-plus-local-build architecture. Background on frontier API pricing: GPT-5.6 API price cut analysis.
Sources: Alibaba Cloud Community blog, Arena.ai leaderboards, Apidog, TechNode, SiliconANGLE, Apple Intelligence China coverage. Verify official pricing, benchmarks, and open-weight status before production decisions.
No. The API is live through Alibaba Cloud Model Studio, but model weights have not been published on Hugging Face or ModelScope. Alibaba's site tags it "Open-Source," but that label describes intent, not a shipped artifact — weights for Qwen3.8-Max and smaller Qwen3.8-27B are promised "next week" with no confirmed date.
No single authoritative head-to-head exists yet. The only independent comparison — a blind-reviewed software architecture task — had Kimi K3 at 83/100 versus Qwen3.8-Max at 80/100. Kimi K3 weights are public with an Artificial Analysis score; Qwen3.8-Max offers lower API pricing and broader native multimodal support. See our Kimi K3 guide.
For the full model, yes — a 2.4T checkpoint is a multi-node datacenter artifact even with 95B active. The API sidesteps that entirely. For local deployment, wait for Qwen3.8-27B or pair with Mac Mini M4 cloud rental as an agent build host.
Treat them as vendor claims, not verified results. Every score comes from Alibaba's own harness, including benchmarks Alibaba created (QwenSWEBench, RecreationBench, and others). No independent platform had reproduced GA numbers as of publication, and the Arena entry remains "Preliminary." Test against your own workload or wait for third-party reproductions.
Qwen already powers generative AI inside Apple Intelligence for users in China, running on-device on recent iPhones — a live example of a Chinese model family becoming embedded infrastructure inside a global consumer platform. Deployment details at our help center.