跳到主要内容

Qwen3.8-Max: The First Open-Weight Max-Class Model (When the Weights Ship)

进阶

On August 3, 2026, Alibaba's Qwen team announced Qwen3.8-Max — a 2.4-trillion-parameter Mixture-of-Experts model with 95B active parameters, a 1M-token context, and benchmark numbers that put it slightly ahead of GPT-5.6 Sol Max on OSWorld-Verified. Alibaba framed the launch as the first Qwen "Max"-class model that will be open-weight — a shift, because prior Max-class Qwens were API-only. Weights were promised for the week of August 10, 2026 on Hugging Face and ModelScope.

There's a catch worth reading before you rewrite anything to depend on it: at launch and for the days after, the Hugging Face page was empty. This page walks the actual state of Qwen3.8-Max as of early August 2026 — the architecture, the pricing, the dual-spec API that makes it a drop-in for Claude or OpenAI clients, the benchmark table with a "vendor-reported" asterisk, and the durable questions to ask before you commit code to it.

What you'll learn
  • Understand what shipped on August 3 vs what was only promised — API access vs open weights
  • Read the 2.4T / 95B active MoE sparsity math correctly, especially before pricing out a self-hosted deployment
  • Know that Qwen3.8-Max exposes both OpenAI-spec AND Anthropic-spec APIs — a rare property that changes portability
  • Interpret the launch benchmarks (OSWorld 86.1, PaperBench 93.0) with the right skepticism — vendor-reported, no model card, no third-party runs yet
  • Recognize the 'open-weight promised ≠ open-weight shipped' pattern and what to do while you wait

The one-sentence version​

Qwen3.8-Max is a 2.4-trillion-parameter sparse MoE (95B active per token) with 1M-token context, native text + vision input, dual OpenAI-spec and Anthropic-spec APIs, priced at $2/$6 per million input/output tokens — announced August 3, 2026 as API-only, with weights and a companion Qwen3.8-27B promised for open release the week of August 10, 2026.

Three things about Qwen3.8-Max that aren't obvious from the headlines​

1. "Open-weight" was announced, not yet shipped​

Alibaba's launch communications and most coverage described Qwen3.8-Max as the first open-weight "Max"-class Qwen. That's a real shift — Qwen3-Max, Qwen3.5-Max, Qwen3.6-Max, and Qwen3.7-Max were all API-only. But at launch, and through the first week, the Qwen organization on Hugging Face had no Qwen3.8-Max repository, no Qwen3.8-27B repository, and no license declared. The commitment was "weights the week of 10 August"; verify before you plan a compliance case around it.

Community reaction in the first hour of launch was almost entirely about the weights, not the benchmarks. Users pushed back on Alibaba's "open-source" framing while the artifacts weren't public. The useful distinction to internalize:

  • An "available" model is a service you rent on the vendor's terms. If the API is deprecated, your dependency breaks.
  • An "open-weight" model is an artifact you download and hold. You can audit it, run it in an air-gapped cluster, keep it after the vendor pivots.

They are different products with different risk profiles. Prior Qwen open-weight releases (Qwen 3.5, Qwen 3.6) shipped under Apache-2.0, which is a strong pattern but not a commitment for Qwen3.8. Read the license the day it lands; don't assume.

2. The 4% sparsity ratio does not mean 4% of the hardware​

Qwen3.8-Max is a Mixture-of-Experts. Of the 2.4T total parameters, only ~95B activate per token — roughly 4% sparsity. That's why the price-per-token can be competitive and the latency reasonable: you're paying for ~95B of forward-pass compute, not 2.4T.

The trap is that weight memory doesn't work that way. To run the model, you need to have all 2.4T parameters resident somewhere fast enough for the router to pull the right experts. In fp8, that's roughly 2.4 TB of weight memory. In fp16, ~4.8 TB. Even with aggressive MXFP4 or similar 4-bit quantization, you're looking at ≥1.2 TB just for weights, before you count KV cache for a 1M-token context. That means:

  • Single H100 (80 GB) — not viable for the full model, even at 4-bit.
  • 8×H100 node (640 GB) — still short at 4-bit for the full 2.4T; possible with heavy CPU offload, at large latency cost.
  • Multi-node inference (16+ H100 or comparable) — the realistic footprint for real-time serving.

If someone quotes you "you can run this on a single machine because it's only 95B active" — they are counting the wrong number. The companion Qwen3.8-27B dense model exists partly for this reason: it's the one you'll actually be able to run on a single 8-GPU node without heroics.

3. Both an OpenAI-compatible and an Anthropic-compatible API — from the same endpoint​

Qwen3.8-Max's API supports both the OpenAI /v1/chat/completions shape and the Anthropic /v1/messages shape. Most vendors ship one; a few ship an OpenAI-compat shim; Qwen shipping first-class Anthropic-spec support is genuinely unusual and worth planning around.

Practical consequence: code you wrote against the Anthropic Python SDK for Claude — including messages.create(), tool blocks, system-prompt handling, and streaming — can point at Qwen3.8-Max with just a base-URL swap and no message-shape rewrites. Same for OpenAI SDK code. This makes A/B testing against Claude or GPT much cheaper than the usual "port your tool schemas" grind.

Qwen3.8-Max via Anthropic-compatible API (Python)

from anthropic import Anthropic

client = Anthropic(
  api_key="YOUR_DASHSCOPE_KEY",
  base_url="https://dashscope.aliyuncs.com/api/v1",  # verify current base URL in Qwen docs
)

response = client.messages.create(
  model="qwen3.8-max",
  max_tokens=1024,
  messages=[{"role": "user", "content": "Plan a migration from Postgres 15 to 17 for a 4TB OLTP database."}],
)
print(response.content[0].text)

Qwen3.8-Max via OpenAI-compatible API (Python)

import openai

client = openai.OpenAI(
  api_key="YOUR_DASHSCOPE_KEY",
  base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
)

response = client.chat.completions.create(
  model="qwen3.8-max",
  messages=[{"role": "user", "content": "Refactor this Python coroutine to use asyncio.TaskGroup."}],
)
print(response.choices[0].message.content)

What the launch benchmarks say — and what the asterisks mean​

Alibaba's launch numbers (self-reported, no model card, no reproducible eval traces at launch):

BenchmarkQwen3.8-MaxGPT-5.6 Sol MaxClaude Fable 5Gemini 3.1 Pro
OSWorld-Verified (agentic OS use)86.183.285.076.2
PaperBench (research synthesis)93.0———
Frontend Code Arena4th place———

The specific asterisks to remember before quoting these to anyone:

  • Vendor-reported. No model card, no eval harness commit, no third-party replication at launch — including from Artificial Analysis, which normally publishes independent scores within days.
  • OSWorld-Verified is a small delta at the top. Qwen 86.1, Claude Fable 5 85.0, GPT-5.6 Sol Max 83.2 is a 1–3 point spread. Within the noise band of many agentic benchmarks. "Beats the frontier by 1.1 points on one benchmark" is not "displaces the frontier."
  • PaperBench is less established than SWE-bench or GPQA. A 93.0 headline is impressive but you should treat it as directional, not definitive.
  • Frontend Code Arena put Qwen3.8-Max fourth — behind several Anthropic models and a Fable 5 variant. Different tasks reward different training signals; a #1 on one benchmark and a #4 on another is a normal outcome, not a contradiction.

Waiting for independent evals (Artificial Analysis, LMArena, third-party academic runs) before rewriting your model-selection default is the sober move.

Pricing math worth knowing​

The API pricing table:

TierPrice per MTok
Input$2.00
Output$6.00
Cached input$0.25

That's 8× cheaper input cost than the current Anthropic Max-tier and about half of GPT-5.6 Sol Max output cost. The cached input at $0.25/MTok is the interesting one: on repeated agentic tool loops with the same system prompt, cache hit rates typically run above 80%, so the effective input cost for a working agent is closer to $0.35–$0.60/MTok, not $2.00.

Compared to the Kimi K3 economics — where cache-hit input is $0.30/MTok — Qwen3.8-Max is in the same low-cache-price band. That makes both viable for cost-sensitive long-horizon agentic workloads where the same prompt scaffolding is re-sent every tool step.

Getting started with Qwen3.8-Max (today, via API)​

Guided walkthrough1 of 4
  1. The Qwen API is served through DashScope (China region) and Model Studio (international). Both expose the qwen3.8-max model ID once you've enabled the model in your workspace and generated an API key.

Qwen3.8-Max vs the models it competes with​

Qwen3.8-MaxKimi K3GLM-5.2Claude Fable 5GPT-5.6 Sol Max
Total params2.4T MoE2.8T MoE753Bundisclosedundisclosed
Active per token~95B16/896 experts———
Context window1M1M—1M+400K+
Native visionYesYesText-onlyYesYes
Open weightsAnnounced (week of Aug 10)Shipped July 27, 2026ShippedNoNo
Input price / MTok$2.00 ($0.25 cached)$3.00 ($0.30 cached)Free (self-host)Vendor tierVendor tier
Output price / MTok$6.00$15.00Free (self-host)Vendor tierVendor tier
API specOpenAI + AnthropicOpenAI-compatVariesAnthropicOpenAI

Reach for Qwen3.8-Max when: you want a frontier-tier API you can also self-host later, your budget is tighter than the Anthropic/OpenAI Max tier, or you want to A/B test against Claude with minimal code changes (thanks to the Anthropic-spec support).

Reach for Kimi K3 when: you need open weights today (K3 shipped July 27, 2026), or you're already on Moonshot tooling.

Reach for GLM-5.2 when: self-hosting is a hard requirement now and vision is not critical.

Stay on Claude Fable 5 or GPT-5.6 Sol Max when: you need the deepest third-party benchmark record, formal enterprise vendor commitments, or the specific tool-use / safety behaviour those models are tuned for.

See also: DeepSeek, Qwen & the Open-Weight Wave for the broader landscape context, and Choosing a model for the general framework.

Practical checklist before you commit to Qwen3.8-Max​

  1. Confirm weights are actually published on Hugging Face or ModelScope before writing "open-weight" into a compliance document.
  2. Read the license on the day it lands. Apache-2.0 is the strong pattern from Qwen 3.5 and 3.6; not a guarantee for 3.8.
  3. Do your own eval run. Vendor OSWorld-Verified numbers are directional; the 1–3 point spread at the top of that benchmark is inside the noise band.
  4. Cost-model the cache carefully. The $0.25/MTok cached-input price is where the economics get interesting; without cache hits the $2/$6 headline is the effective price.
  5. Plan the self-hosting footprint honestly. 4% active parameters ≠ 4% GPU memory. Full-weight residency for a 2.4T MoE is a serious deployment, not a laptop project.
  6. Decide which spec you'll code against. OpenAI-spec and Anthropic-spec are both first-class — pick whichever matches your existing code path, don't invent a third abstraction.

Quiz​

Check yourself

0/5
  1. Qwen3.8-Max was announced on August 3, 2026 as 'open-weight'. What was actually available at launch?
  2. Qwen3.8-Max is 2.4T total parameters with 95B active per token. What does this imply for self-hosting hardware?
  3. What's unusual about Qwen3.8-Max's API compared to most competing models?
  4. Alibaba's launch showed Qwen3.8-Max at 86.1 on OSWorld-Verified vs GPT-5.6 Sol Max at 83.2. What's the right way to read that?
  5. Qwen3.8-Max lists $2/$6 per million input/output tokens, with cached input at $0.25/MTok. For a long-running agentic loop with a large stable system prompt, what's the practical effective input cost?
按 Enter 或空格键翻转卡片。使用左右方向键在卡片之间切换。已显示术语。
1 / 8

Sources & further reading​

Next​