Glonce

Qwen3.8-Max arrives as Qwen's first open-weight Max-class model

modelsThe Glonce editorsHacker News🔥80Glonce rating80Impact65Novelty60Relevance100Surprise45Players70Credibility85

Qwen announced Qwen3.8-Max on August 3, 2026, describing it as the most capable model in the Qwen family to date and the first Qwen-Max-class model to have its weights open-sourced. Built on the Qwen 3.5 architecture, it scales to 2.4 trillion total parameters with 95 billion active. The model is available now through QwenCloud, and open weights are set to follow next week on Hugging Face and ModelScope.

To demonstrate autonomous coding, Qwen3.8-Max was asked to build the oh-my-cli project from scratch and, over a 10-plus day autonomous run with no human help, construct a self-evolving development harness combining an issue state machine, dispatcher, monitor and watchdog, plus automated build, unit, end-to-end and desktop lifecycle testing. As of July 30, 2026, about 16 days into fully autonomous operation, the public repository (qwen-code-dev-bot/oh-my-cli on GitHub) had accumulated 265 commits, 127 pull requests and 151 issues.

In a second test, the model was handed the paper "Unified Data Selection for LLM Reasoning" with no starter code and asked to reproduce it, then improve on it. Working alone for about five days (roughly 125 hours), it spent the first 37 hours rebuilding the paper's full pipeline from nothing and reproducing its six main findings, confirming that the paper's method for selecting hard training examples beats random selection by 7.7 percentage points on the AIME24 math benchmark when fine-tuning a Qwen3-8B model. It then spent about 88 more hours running its own hypothesize-code-test-analyze loop, trying 18 self-generated improvement ideas across four rounds, and ultimately produced a new method that beat the paper's own approach by 2.7 points on AIME24. In total it wrote roughly 7,600 lines of code, took over 1,100 actions and ran 33 rounds of GPU training.

Qwen3.8-Max was also entered into the WWW2025 Multimodal Dialogue Intent Recognition Challenge on Alibaba Cloud's Tianchi platform, competing against 526 human teams to classify customer-service chats from text and screenshots. Working alone under a 24-hour limit, it fine-tuned and ensembled BERT, MacBERT and RoBERTa for text, added a fine-tuned Qwen2.5-VL-7B backed by Chinese-CLIP for images, and combined everything through weighted voting calibrated by cross-validation. Across 45 submissions its accuracy climbed from 0.60 to 0.853, beating 458 of the 526 teams, or 87 percent of the field.

Qwen also stress-tested the model's breadth across several hundred high-value professions. Examples cited: a compliance-counsel task surfaced 1,284 relevant clauses across hundreds of documents in under an hour, versus roughly a week for a paralegal team; a UI/UX design task produced an 8-screen prototype for a banking app called NOVA in one pass with zero revision rounds, versus 3 to 5 rounds typically; a restaurant-menu task read over a hundred ingredient-supply briefs and produced a 26-dish menu with food cost held at 33.8 percent; a structural-engineering task reconstructed the seismic model of a 30-story tower from drawings, work that normally takes an engineer over a week in specialized software; a rehabilitation task turned a 2D paper assessment into a rotatable 3D anatomical demo, replacing outsourced medical-animation work that costs thousands of dollars and takes 2 to 4 weeks; a sports-analytics task parsed about 8,400 offensive and defensive possessions per player into a tactical report in tens of minutes, versus several working days for a human analytics team.

Given a one-line brief, the model also built a complete ETF-rotation quant strategy end to end, including its own data system, factor construction and iterative overfitting correction. In a separate, parallel test it took six short descriptions of classic factor families, decomposed them into 50 research directions each, dispatched about 330 sub-agents and ran about 6,000 backtests; the resulting factors reached excess Sharpe ratios of 0.64 to 1.48 with information coefficients between 0.010 and 0.014.

For a long-horizon hardware test, Qwen3.8-Max autonomously designed a GCD/RSA cryptographic accelerator chip using only a stub RTL workspace and an evaluation script, with no reference design and no human intervention, working through simulation (Iverilog), synthesis (Yosys) and physical layout (OpenROAD) toolchains. Over about 500 turns, 71 evaluations and 13 milestones, it cut the gate count from an initial working design of 8,298 gates down to 678, the lowest among evaluated models; the single largest step replaced a hardware modulo divider with an iterative shift-subtract design, removing 6,288 gates in one move. Carried through to physical layout, the chip's die shrank from 106 by 106 micrometers (33,369 micrometers of wirelength, with a timing violation of -4.46 nanoseconds slack) to 46 by 46 micrometers (4,187 micrometers of wirelength), an 81 percent area reduction, while achieving timing closure at 500 MHz.

In E-Commerce Bench, a 365-day simulated retail-operations benchmark built on desensitized Taobao and Tmall data spanning 12 store types, 60 product categories, nearly 600 suppliers (152 of them deliberately fraudulent) and 7,000 products, Qwen3.8-Max started with 100,000 yuan and had to handle sourcing, price negotiation, inventory, pricing and returns while converting gains to cash by year end. It improved its procurement prices round over round through repeated negotiation, invested heavily early to build position, earned over 100,000 yuan in net profit during the year-end promotion period alone, nearly 2.4 times the runner-up, and finished with a total balance of 416,252 yuan, a 4.16x return that beat second-place GLM 5.2 by 38 percent and improved 152 percent on its own predecessor, Qwen3.7-Max.

On the multimodal side, Qwen says the model can work through financial reports and PDFs over 200 pages long and process videos longer than 100 hours, building a searchable "video memory graph" of people, events, timestamps and scenes. It also performs visual production tasks such as editing footage into a vlog, turning a floor plan into a 3D interior in Blender, or rebuilding a frontend from a single screenshot, and checks its own intermediate visual output during a task to catch and correct mistakes. On RecreationBench, a new benchmark spanning desktop, mobile and web, the model rebuilds a real running application it can only observe as a black box, with no source code or internet access. A new extension library, Qwen-MM-Plugins, is meant to bring these multimodal capabilities into existing agent harnesses.

The model supports a reasoning_effort parameter (xhigh by default, plus medium and low settings) to trade reasoning depth against speed and cost, with preserve_thinking enabled by default, and Qwen says it performs comparably across harnesses including QwenWork, Claude Code, Codex, OpenClaw and Hermes.

Key facts

  • Qwen3.8-Max scales to 2.4 trillion total parameters (95 billion active), built on the Qwen 3.5 architecture; it is the first Max-class Qwen model whose weights will be open-sourced, due next week on Hugging Face and ModelScope.
  • In a 10-plus day autonomous coding run building the oh-my-cli project, the model's self-evolving harness had produced 265 commits, 127 pull requests and 151 issues by July 30, 2026, about 16 days in.
  • Working alone for about 125 hours, it reproduced a research paper's data-selection method (a +7.7 percentage point AIME24 gain over random selection), then spent 88 more hours self-improving on it for a further +2.7-point AIME24 gain.
  • In a 24-hour online contest against 526 human teams, its accuracy climbed from 0.60 to 0.853 across 45 submissions, beating 458 of the 526 teams (87 percent of the field).
  • In a year-long e-commerce simulation it turned 100,000 yuan into 416,252 yuan (a 4.16x return), beating second-place GLM 5.2 by 38 percent and its own predecessor Qwen3.7-Max by 152 percent.

Why it matters

Qwen3.8-Max is pitched as Qwen's most capable model yet and, at 2.4 trillion parameters (95 billion active) built on the Qwen 3.5 architecture, the first Max-class Qwen model to have its weights released openly rather than kept API-only. The launch leans heavily on long-horizon autonomy: multi-day and multi-hour case studies (a 10-plus day coding run, a roughly five-day research reproduction, a 500-turn chip-design session, a year-long simulated business) rather than single-prompt benchmark scores, arguing the model can carry a goal across days without a human resetting or steering it.

Who it affects

Developers using coding assistants and agent harnesses, since Qwen reports comparable performance across QwenWork, Claude Code, Codex, OpenClaw and Hermes, and ships a Qwen-MM-Plugins library for adding multimodal capability to existing harnesses. The professions used as case studies, including compliance counsel, UI/UX designers, restaurant operators, structural engineers, rehabilitation therapists, sports analysts and quant researchers, are presented as workflows the model can substantially speed up. The open-source community is affected once weights land next week on Hugging Face and ModelScope, since this is the first Max-class Qwen release to be open-weighted at all.

How to use it

Qwen3.8-Max is available now via the QwenCloud API. It supports a reasoning_effort parameter with three settings, xhigh by default for thorough analysis, medium to balance accuracy and speed, and low for efficient, cost-optimized reasoning, and ships with preserve_thinking enabled by default. Open weights follow next week on Hugging Face and ModelScope. The announcement does not give token pricing anywhere, so cost per request is not yet public.

How solid is it

The evidence in the post is a set of self-run case studies (the oh-my-cli repository, the research-paper reproduction, the WWW2025 contest, the chip-design run, E-Commerce Bench, RecreationBench) with concrete, specific numbers rather than vague claims, and several of them, like the public GitHub repository and the WWW2025 leaderboard, are independently checkable. That said, every case study was designed, run and reported by the Qwen team itself.

Risks and caveats

All figures come from Qwen's own announcement; none of the case studies has independent, third-party replication cited in the post. No token pricing is disclosed. The open-weight release is dated only as "next week" relative to the August 3, 2026 post, with no exact calendar date given. The article's byline credits only "QwenTeam," without naming individual researchers. No mechanism or timeline is given for how third-party harnesses like Claude Code or Codex would add native Qwen3.8-Max support beyond manual API integration.