Alibaba releases Qwen3.8-Max with 2.4T MoE and 1M context

• Alibaba shipped Qwen3.8-Max as its biggest Qwen flagship so far: 2.4T total parameters, 1M context, native text and vision, and an API launch through Alibaba Cloud Model Studio, according to hasantoxr's release post.
• The serving story is the sparse MoE: only 95B parameters activate per token, a detail hasantoxr's architecture note called more important than the 2.4T headline.
• The flashiest demo was a long agent run: Qwen3.8-Max built oh-my-cli over 16 days by turning requests into GitHub issues, assigning itself work, writing code, running tests, and shipping fixes, as hasantoxr's coding summary described it.
• Pricing landed at $2 per 1M input tokens, $6 per 1M output tokens, and $0.25 cached input, with open weights promised for the following week in hasantoxr's pricing post.
• Creative hands-on reports already pointed past coding: illscience used Qwen3.8-Max in an all-day agent workflow to script, mix, cut, and direct a five-minute documentary, while kaigani tested it on video continuity errors.

Alibaba's coding case study includes the full oh-my-cli loop, plus a research-paper reproduction run that wrote about 7,600 lines of code and ran 33 GPU training rounds. The public oh-my-cli repo calls itself “a minimal autonomous code-agent CLI,” runs on Node.js 22 and TypeScript, and ships under Apache-2.0. A Capital & Compute benchmark read found the awkward bit: cheaper per-token pricing did not mean cheaper completed work on Artificial Analysis's full Intelligence Index run.

Alibaba's official launch post describes Qwen3.8-Max as the most powerful model in the Qwen series to date, with API access through Alibaba Cloud Model Studio and weights scheduled for release the following week.

• Model: Qwen3.8-Max, built on Qwen 3.5.
• Architecture: sparse mixture-of-experts with hybrid attention.
• Scale: 2.4T total parameters, 95B active parameters.
• Context: up to 1M tokens.
• Modalities: text plus native visual intelligence for images, documents, video, and UI screenshots.
• API: available through Alibaba Cloud Model Studio and QwenCloud.
• Price: $2 input, $6 output, $0.25 cached input per 1M tokens, per hasantoxr's pricing post.
• Open weights: promised for the following week, the first Qwen-Max-class weight release Alibaba has announced.

The real creator hook is the input budget. Alibaba says the model can ingest hundred-page documents, full TV series, and 100-hour livestreams, then turn them into searchable knowledge structures.

First-party

• DeepSWE 1.1: 21.6 → 56.6, +35.0 points, per hasantoxr's benchmark excerpt.
• FrontierSWE: 40.7 → 73.5, +32.8 points, per hasantoxr's benchmark excerpt.
• JobBench: 31.3 → 53.4, +22.1 points, per hasantoxr's benchmark excerpt.
• WWW2025 challenge accuracy: 0.60 → 0.853 across 45 submissions, +25.3 points, according to Alibaba's coding case study.

Third-party evaluators

• Artificial Analysis Intelligence Index: Qwen3.7 Max 47 → Qwen3.8 Max 58, +11 points, according to Capital & Compute's same-day benchmark read.

Customer-reported

• No scored customer benchmark deltas surfaced in the tweet pool. The hands-on reports below describe creative workflows rather than eval tables.

Alibaba also listed OSWorld-Verified at 86.1, Parametric CAD Bench at 91.5, and OmniDocBench at 92.1 in the launch thread, but the evidence pool did not include predecessor values for those scores.

The per-token cut hid a cost-per-task gotcha.

Capital & Compute's benchmark read says Alibaba cut list rates by 20% versus Qwen3.7 Max, but Artificial Analysis spent $1,063.86 evaluating Qwen3.7 Max and $1,741.41 evaluating Qwen3.8 Max on the same Intelligence Index suite, +64%.

The same read measured output speed at 201.9 tokens/sec for Qwen3.7 Max and 67.6 tokens/sec for Qwen3.8 Max, -134.3 tokens/sec. For agentic runs with dozens of calls, that turns a token discount into a wall-clock tradeoff.

Alibaba's launch post says Qwen3.8-Max combines sparse MoE routing with hybrid attention, activating 95B of 2.4T parameters during inference.

The model is pitched as a multimodal foundation, not a text model with a vision adapter stapled on. hasantoxr's multimodal note says it can take screenshots, PDFs, and long-form video, while illscience's reply called it “natively multimodal.”

The same official post names visual coding tasks that matter for designers: reconstructing frontend projects from UI screenshots, turning 2D floor plans into 3D interior visualizations, editing raw footage into vlogs, and building interactive games from natural-language requests.

Illscience left Qwen3.8-Max running all day to generate a five-minute National Geographic-style documentary about Hall Beach, Nunavut. The workflow used OpenCode, MiniMax, FAL, and ElevenLabs, and the model wrote the script, mixed audio, made clips, and “directed” the video.

The follow-up is the useful creative note: illscience said the test was aimed at long-horizon capability, but the surprising part was storytelling quality compared with other models.

Kaigani tested Qwen3.8 on video understanding by asking it to find continuity errors. In a separate reply, kaigani said they wanted to try Qwen3.8 for video QC once a local version exists, after Claude missed a floating head at the end of a local CODEYWOOD sci-fi short.

• Alibaba Cloud Model Studio: the model is available by API, according to Alibaba's official launch post.
• QwenCloud: Alibaba's coding post says Qwen3.8-Max is callable through the QwenCloud API.
• Qwen Studio: the Qwen Studio link in the launch thread points to the consumer-facing Qwen interface.
• QwenWork: Alibaba's QwenWork announcement says the workplace agent platform is in public beta in China, with web access, a desktop client, and planned DingTalk embedding.
• Open weights: Capital & Compute's August 7 check found no Qwen3.8 repo or named license yet, while hasantoxr's pricing post said the weights were scheduled to land on Hugging Face the following week.