guide · ai

Qwen3.8-Max Weights: Out

Qwen3.8-Max Weights: Out (Alibaba / Qwen, August 11-12, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.

August 12, 2026 · By Alastair Fraser

A retro robot representing Qwen3.8-Max Weights: Out

title: “Qwen3.8-Max Weights: Out” description: “Qwen3.8-Max Weights: Out (Alibaba / Qwen, August 11-12, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.” type: guide category: ai pubDate: 2026-08-12 image: /images/abs-model-qwen3.8-max-weights-update.png imageAlt: “A retro robot representing Qwen3.8-Max Weights: Out” imagePrompt: “Bold graphic editorial illustration, 1990s comic-book influence, heavy ink outlines, halftone texture, crimson and electric blue on cream. A single retro-futurist robot representing an AI model / a brain-in-a-server, no text, no logos, 16:9.” affiliate: false sources: facts:

label: “Vendor” value: “Alibaba / Qwen”
label: “Released” value: “August 11-12, 2026”
label: “License” value: “Open-weight (custom Qwen3.8-Max License)”
label: “Type” value: “AI model release” related:
abs-model-claude-opus-5
abs-model-gpt-5.6-cyber
abs-model-grok-4.6
abs-model-kimi-k3
abs-model-lfm2.5-vl-3b
abs-model-ling-3.0-flash
abs-model-muse-glimmer-30b
abs-model-muse-spark-1.2
abs-model-nemotron-3.5-lightning
abs-model-qwen-image-3.0
abs-model-qwen3.7-flash
abs-model-qwen3.8-2.4t-a95b
abs-model-solar-pro-4 tags:
qwen3.8-max-weights-update
ai-models
model-release draft: false

label: “Vendor” value: “Alibaba / Qwen”
label: “Released” value: “August 11-12, 2026”
label: “License” value: “Open-weight (custom Qwen3.8-Max License)”
label: “Type” value: “AI model release” related:
abs-model-claude-opus-5
abs-model-gpt-5.6-cyber
abs-model-grok-4.6
abs-model-kimi-k3
abs-model-lfm2.5-vl-3b
abs-model-ling-3.0-flash
abs-model-muse-glimmer-30b
abs-model-muse-spark-1.2
abs-model-nemotron-3.5-lightning
abs-model-qwen-image-3.0
abs-model-qwen3.7-flash
abs-model-qwen3.8-2.4t-a95b
abs-model-solar-pro-4 tags:
qwen3.8-max-weights-update
ai-models
model-release draft: false

The short version

Yes — the weights are out. Qwen3.8-Max landed on Hugging Face and ModelScope on Aug 11–12, 2026 as Qwen3.8-2.4T-A95B (BF16 + FP8). It’s a real open-weight release of a genuine frontier-tier MoE: 2.4T total parameters, 95B active. This is the first time a Qwen Max tier has gone open, and it’s the largest open-weight drop by raw bytes shipped (~4.9 TB BF16). But three things separate this from the “Apache 2.0, download and go” releases you’re used to: a custom license with commercial triggers, a slow API-to-weights cadence, and a size that makes self-hosting a rented-GPU decision, not a Mac-mini one.

What actually shipped

Hugging Face primary: Qwen/Qwen3.8-2.4T-A95B (BF16) plus an …-FP8 variant (~2.5 TB).
ModelScope mirror for CN access.
OpenRouter listed the open-weight variant Aug 12; Fireworks AI announced day-0 serving the same day.
Compatible with Transformers, vLLM, SGLang, and TokenSpeed. No training data, no distillation code, no post-training recipes — weights, config, inference code only.

The API itself launched earlier: previewed Jul 19, 2026 at WAIC Shanghai, officially launched Aug 2 on Qwen Cloud at $2/$6 per 1M tokens. The model card is text-only, 262K native context (extensible to 1M via YaRN), and thinking mode is always on.

The license is not Apache — read this

This is the part people get wrong. The HF badge may render like a permissive license, but the LICENSE file governs, and it is a custom Qwen3.8-Max License — not Apache 2.0. Do not describe this as “Apache 2.0 open-source.”

The core is MIT-permissive: free use, copy, modify, merge, sell, deploy, fine-tune, derive. Two commercial triggers sit on top:

Attribution required above 100M MAU or $20M monthly revenue (model name must be prominent).
Separate paid license required above $50M annual revenue for “Model as a Service” or “AI Work Assistant” businesses.

ABS is far below every threshold, and there are no geographic restrictions — so for us it’s effectively open weights on a custom permissive-ish license. Still, document the terms in your provider profile; the MaaS clause matters the day a consumer brand crosses $50M/year.

How the cadence compares

The API-to-weights gap was 23–24 days from the Jul 19 preview (9–10 days from the Aug 2 official launch). Against Moonshot’s Kimi K3 — 11 days (Jul 16 API → Jul 27 weights) — that’s roughly 2× slower. Alibaba matched K3’s “open-with-a-toll-booth” template but at a lower tempo. Two readings: either aligned intent with slower execution (a 2.4T model plus a bespoke license costs calendar time), or a deliberate “proprietary wrapper first” policy — the API ships vision, 1M-context default, and built-in tools the weights don’t have. Either way, the 24-day gap is a policy shift, not a slip: prior Qwen Max tiers stayed closed, and smaller Qwen3.5/3.6 models shipped same-week.

What’s missing from the weights

No vision (use Qwen3.7 Flash for cheap multimodal), no 1M context by default, no built-in tool harness, thinking mode can’t be disabled, and no QAT-4bit at launch (BF16 + FP8 only). Unsloth’s 1-bit dynamic quant (~397 GB) is the only single-node path, and 95B active params per forward pass still demands serious accelerators. Practical floor for BF16: 8× H100/B200.

What it means for you

Use the API today. There’s no cost advantage to self-hosting at our scale — same $2/$6 pricing on Fireworks/OpenRouter.
Watch the 27B sibling (~Aug 14, 2026). Same post-training, runs on a single Mac / DGX Spark. Treat A95B as the strategic roadmap; 27B is the daily-driver candidate.
Don’t call it Apache. Open weights on a custom license with revenue/MAU thresholds — cite the HF LICENSE file, not the badge.