Qwen 3.8 Max Ships: 2.4T MoE, 1M Context, $2/$6 per MTok, Open Weights Next Week

Alibaba released Qwen 3.8 Max on August 3, 2026: a 2.4-trillion-parameter MoE with 95B active per token, a 1M-token context window, and native text plus vision input. QwenCloud prices it at $2 per million input tokens and $6 per million output, and the weights - the first Max-class model Alibaba has ever open-sourced - are promised for next week on Hugging Face and ModelScope.

The model is live on two surfaces today: QwenCloud under the model id qwen3.8-max, and Vercel's AI Gateway as alibaba/qwen3.8-max (added August 2, provider pricing, no markup). The teaser tweet from July 19 said "launching and going open-weight soon"; the launch is now real, and the open-weights promise is the headline for anyone running self-hosted fleets.

Official Sources#

What Shipped#

Qwen 3.8 Max is built on the Qwen 3.5 architectural foundation and scales it up. The concrete numbers from the announcement:

• 2.4T total parameters, 95B active per token. That places it between Moonshot's Kimi K3 (2.8T total, 104B active) and the previous Qwen 3.7 Max.
• 1M context window, with 991K max input and 131K max output. Reasoning chains can extend to 262K tokens.
• Text, image, and video input in one model. Qwen also shipped Qwen-MM-Plugins, a harness extension library for multimodal agents (video memory, dynamic resolution, visual tool use).
• reasoning_effort control: xhigh (default), medium, and low levels, plus preserve_thinking on by default. This is the same cost-control pattern DeepSeek and others now use - drop the effort level for cheap fast passes, raise it for long-horizon work.
• Open weights next week. The first Max-class release in Qwen history, destined for Hugging Face and ModelScope. At 2.4T parameters this is a datacenter-scale download (K3 weighed in around 1.63TB), but it changes what "frontier" means for self-hosted operators.

Benchmarks#

All numbers below are vendor-published in the announcement, with the caveats Qwen states: most external models were evaluated on their own preferred harnesses (Claude Code for the Claude models, Codex for GPT-5.6 Sol), and some Qwen in-house benchmarks are new. Treat cross-lab comparisons as directional.

The shape is familiar from K3 and GLM 5.2 launches: the new open-weights flagship beats the incumbent closed frontier on some agentic and multimodal surfaces (PaperBench, OSWorld-Verified, RealWorldQA) while trailing on others (DeepSWE, Fable 5's FrontierSWE). One note for this site specifically: Qwen states the Qwen series was evaluated on SkillsBench using OpenCode as the harness - the model runs in our own tooling.

Two showcase runs from the announcement are worth more than the table. In a Tianchi competition with 526 human teams, Qwen 3.8 Max worked autonomously for 24 hours, submitted 45 times, and climbed from 0.60 to 0.853 accuracy, beating 458 of 526 teams (87%). And handed a research paper plus a GPU budget, it reproduced the paper's pipeline from zero (7,600 lines of code, 33 training rounds), then improved on the paper's own method by +2.71 points on AIME24 over four self-designed experiment rounds. These are marketing demos, but they demonstrate the long-horizon loop that the benchmarks measure indirectly.

Pricing#

Live from the QwenCloud model page, verified August 3, 2026:

Context caching is the feature that changes the effective cost. At $0.25/MTok on implicit cache reads, an agent loop that re-reads the same 200K-token repository across 50 turns pays about 8x less on input than the uncached rate - the same cache-first economics we tracked in the frontier pricing tracker.

For comparison: Qwen 3.7 Max sits at $1.25/$3.75, Kimi K3 at $3.00/$15.00 on cache miss, and DeepSeek V4 Flash at $0.14/$0.28. Qwen 3.8 Max is the premium tier of its own family - 60% above 3.7 Max on input - but still well under K3's output rate.

Running It Today#

The model is not yet in OpenCode's model registry (the registry currently carries the Qwen 3.5 Plus, 3.6 Plus, and 3.7 Max lines). Until it lands, the fastest path is the QwenCloud API, which speaks both the OpenAI and Anthropic protocols. The official quickstart shows the OpenAI-compatible route:

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ.get("DASHSCOPE_API_KEY"),
    base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
)
response = client.chat.completions.create(
    model="qwen3.8-max",
    messages=[{"role": "user", "content": "Explain this repo's build failure"}],
    extra_body={"enable_thinking": True},
    reasoning_effort="xhigh",
    stream=True,
)

Because Qwen exposes the Anthropic protocol at https://dashscope-intl.aliyuncs.com/apps/anthropic, you can also point Claude Code at it directly (ANTHROPIC_MODEL=qwen3.8-max with that base URL), and the Codex model catalog ships a qwen3.8-max entry with the same reasoning levels. The Vercel AI Gateway route is the zero-markup option with usage tracking, and vercel ai-gateway coding-agents setup wires it into Claude Code, Codex, or OpenCode once the model lands there.

Decision Guide#

• Agentic coding on a 1M context budget: Qwen 3.8 Max at $2/$6 is the cheapest way to get K3-class long-context autonomy from a frontier-adjacent model, and cache reads make repeated repo scans cheap.
• Self-hosting or fine-tuning: wait for the open weights next week. 2.4T total is not consumer hardware, but it is deployable on the same clusters that run K3 or GLM 5.2, with 95B active for inference throughput.
• Cost-sensitive volume work: DeepSeek V4 Flash at $0.14/$0.28 is still a third of the input price and a twentieth of output. Qwen 3.8 Max earns its premium on long-horizon and multimodal tasks, not on quick edits.
• Multimodal agents: the vision-plus-execution loop (observe, verify, correct) plus Qwen-MM-Plugins makes it the strongest open-weights multimodal agent option on paper today.

The open-weights week is the event to watch. If the license is permissive like GLM 5.2's, the self-hosted frontier gets a 2.4T option that undercuts every closed lab on the agentic surfaces where it leads.

FAQ#

What is Qwen 3.8 Max?#

Alibaba's flagship MoE released August 3, 2026: 2.4T total parameters, 95B active, 1M context, text plus vision input, priced at $2/$6 per million tokens on QwenCloud. Weights are set to open on Hugging Face and ModelScope the following week.

What does Qwen 3.8 Max cost?#

$2.00 per million input tokens and $6.00 per million output, with implicit cache reads at $0.25/MTok, explicit cache creation at $2.50/MTok, and explicit cache reads at $0.17/MTok. Vercel AI Gateway mirrors provider pricing with no markup.

Is Qwen 3.8 Max available in OpenCode?#

Not yet - the OpenCode model registry currently lists Qwen 3.5 Plus, 3.6 Plus, and 3.7 Max. Qwen's own SkillsBench runs used OpenCode as the harness, so the wiring is a model entry away. In the meantime, use the QwenCloud API directly or through Claude Code and Codex via the Anthropic/OpenAI-compatible endpoints.

What is the context window?#

1M tokens: 991K max input, 131K max output, and reasoning chains up to 262K tokens.

When do the open weights release?#

Qwen says next week, on Hugging Face and ModelScope. It will be the first open-weights release of a Max-class Qwen model.

Sources#

Last updated: August 3, 2026

Continue Reading#

• Qwen 3.7 Max Developer Guide - the previous generation flagship: 1M context, $1.25/$3.75, agent-first architecture
• Kimi K3: Moonshot's 2.8T Frontier Model - the 2.8T open-weights rival Qwen 3.8 Max is priced against
• GLM 5.2 vs DeepSeek V4 vs Qwen3: Open-Weights Coding Showdown - where the open frontier stands across the three families
• DeepSeek V4 Flash 0731: OpenCode Guide - the low-cost end of the same market at $0.14/$0.28
• Frontier Model API Pricing, June 2026 - the running tracker these price points extend