Qwen3.8-Max Release: 2.4T Parameter Model with Open Weights

Qwen3.8-Max Overview

Qwen3.8-Max is the most capable model in the Qwen family to date, released on August 3 2026. It scales to 2.4 trillion parameters with 95 billion active parameters and is the first Qwen‑Max class model whose weights will be open‑sourced (scheduled for release the following week). Built on the architectural foundation of Qwen 3.5, it delivers comprehensive improvements across coding, work, research, and long‑horizon tasks, enabling end‑to‑end completion of complex goals with minimal human involvement.

Coding Capabilities

Qwen3.8-Max can autonomously conduct multi‑day software projects from scratch, using self‑evolving feedback loops. In an oh‑my‑cli harness experiment, the model ran for over 10 days (approximately 16 days total) without human intervention, producing 265 commits, 127 pull requests, and 151 issues by July 30 2026. The harness normalizes requirements into GitHub Issues, dispatches agents through a state machine, runs build/unit/E2E/desktop lifecycle tests, and iterates based on self‑test results and community feedback.

When given a recent research paper (“Unified Data Selection for LLM Reasoning”) and a set of GPUs, Qwen3.8‑Max reproduced the paper’s pipeline from scratch in about five days (~125 hours), wrote roughly 7,600 lines of code, executed 33 GPU training runs, and confirmed the paper’s +7.7% gain on AIME24 over random data selection. It then entered a self‑improving loop, testing 18 ideas across four rounds and arriving at a new method that outperforms the paper’s approach by +2.7 points on AIME24.

In the WWW2025 Multimodal Dialogue Intent Recognition Challenge on Alibaba Cloud’s Tianchi platform, Qwen3.8‑Max operated under a strict 24‑hour limit, built a solution that fine‑tuned BERT, MacBERT, RoBERTa for text and Qwen2.5‑VL‑7B backed by Chinese‑CLIP for images, fused them via a weighted‑voting system, and improved accuracy from 0.60 to 0.853 over 45 submissions. This score beat 458 of the 526 competing human teams (87% of the field).

Work‑Related Competence

Qwen3.8‑Max is designed to handle messy, multi‑step, tool‑heavy workflows across many professions. Scaling real‑world reinforcement learning environments jointly with compute yields broad, reliable gains. The model uses a universal reward system that unifies execution‑based checks, rubric‑conditioned adjudication over text and visual output, and agentic inspection, while an online data balancer keeps task, difficulty, workspace, and harness distributions stable to suppress gradient variance.

Representative showcases include:

  • Corporate compliance counsel: reviewed hundreds of documents scanned in under an hour, surfacing 1,284 relevant clauses (a task that normally takes a paralegal team about a week).
  • UI/UX designer: produced an eight‑screen, high‑fidelity interactive prototype for the NOVA banking app in one shot with zero human revision cycles (versus 3–5 rounds conventionally).
  • Restaurant brand founder: generated a 26‑dish menu with caloric values and ingredient provenance, keeping food‑cost ratio at 33.8% in a single pass (normally weeks of iterative work).
  • Structural engineer: reconstructed a 30‑story office tower’s seismic model in the browser, providing natural period, base shear, and inter‑story drift ratio on hover (traditionally a week‑long manual effort).
  • Rehabilitation therapist: turned a 2D paper assessment into a 3D interactive demo with anatomical overlays, a task previously outsourced to a medical‑animation studio for 2–4 weeks and thousands of dollars.
  • Sports data analyst: parsed ~8,400 offensive/defensive possessions per player into tactical profiles and coaching reports in tens of minutes (versus several days for a traditional team).

Powered by Dynamic Workflows, Qwen3.8‑Max can also devise a profitable end‑to‑end quant strategy in a single conversation. It planned an ETF‑rotation strategy, pruned overfitting factors, added multi‑seed union validation, and switched ensembling strategies based on validation evidence. In parallel factor mining, it decomposed six factor families into 50 into ~3333000000 research directions, dispatched ~330 sub‑agents, ran ~6,000 backtests, and selected factors with excess Sharpe ratios of 0.64–1.48 and IC ranging from 0.010 to 0.014.

Long‑Horizon Task Performance

Qwen3.8‑Max exhibits strong closed‑loop optimization in long‑horizon, multi‑constraint tasks. In an autonomous chip design benchmark for a GCD/RSA cryptographic hardware accelerator, the model started from a blank RTL workspace and, over ~500 turns and 71 evaluations across 13 milestones, reduced the gate count from 8,298 to 678 (an 81% reduction) while achieving timing closure at 500 MHz (+0.66 ns slack) after place‑and‑route with OpenROAD. Key steps included replacing a 16‑bit hardware modulo divider with an iterative shift‑subtract architecture (saving 6,288 gates), redundancy elimination, register and FSM pruning, module fusion, and gate‑level refinements.

In the E‑Commerce Bench (a 365‑day simulated Taobao/Tmall operation), Qwen3.8‑Max began with ¥100,000 capital, negotiated with a supplier matrix containing 152 fraudulent merchants, managed inventory, dynamic pricing, and returns handling. It achieved a net profit exceeding ¥100,000 during the year‑end promotion period—about 2.4 times that of the second‑place GLM 5.2—and finished with a total balance of ¥416,252, a 4.16× return and a 152% improvement over its predecessor Qwen3.7‑Max. The model continuously improved negotiation efficiency over time and generalized learned strategies to similar products.

Multimodal Agent Abilities

Qwen3.8‑Max treats vision as a native feedback loop that spans planning, execution, verification, and iteration. It can understand and generate content across images, documents, and videos longer than 100 hours, building video memory graphs that connect events, timestamps, and scenes. During execution it continuously observes its own intermediate results, detects mismatches (e.g., a misaligned interface or an object facing the wrong direction), revises its plan, and corrects the output autonomously.

The Hybrid Agent capability pairs coding with GUI operation: code handles heavy‑logic scaling while GUI interaction supplies real‑world feedback, extending the visual loop to verify against a live application. The RecreationBench benchmark (spanning Ubuntu, macOS, Windows, Android, and the web) measures this ability; Qwen3.8‑Max achieves a score of 51.7, demonstrating frontier‑level application recreation from a black‑box observation.

To simplify integration, the lab releases Qwen‑MM‑Plugins, a harness extension library providing image/video processing, multimodal memory, dynamic‑resolution support, visual tool use, and specialized capabilities for video editing, Blender, and CAD. Any existing agent framework can be extended with these plugins to become more multimodal‑native.

Benchmark Highlights

Across a suite of internal and public benchmarks, Qwen3.8‑Max shows strong gains over prior Qwen versions and competitive models. Notable results include:

  • Coding Agent: Terminal Bench 2.1 86.6 (vs. 74.5 for Qwen3.7‑Max), SWE‑bench Pro 67.7 (vs. 60.6), DeepSWE 1.1 56.6 (vs. 21.6), PaperBench 93.0 (vs. 64.8), QwenSWEBench 80.7 (vs. 63.4).
  • General Agent: CoWorkBench 74.8 (vs. 64.6), WorkSpaceBench 67.7 (vs. 61.4), JobBench 53.4 (vs. 31.3), SkillsBench 70.2 (vs. 61.2), Agents’ Last Exam 27.0/52.4 (vs. 11.8/31.1).
  • General Capabilities: GPQA Diamond 92.6 (vs. 92.4), IFBench 82.8 (vs. 79.1), HealthBench 60.2 (vs. 54.5), PLawBench 73.2 (vs. 58.9), PRBench‑Finance 58.3 (vs. 46.8).
  • Multimodal Reasoning: MMMU‑Pro 82.3 (vs. 79.0 for Qwen3.7‑Plus), MathVision 95.2/97.7 (vs. 90.3/--), BabyVision 82.0/91.3 (vs. 64.7/70.4), LogicVista 91.9 (vs. 84.3), HiPhO 90.0 (vs. 84.1), PhyX 83.5 (vs. 80.0), SLAKE 90.8 (vs. 83.2).
  • Visual Agent & Coding: OSWorld‑Verified 86.1 (vs. 73.3), AndroidWorld 85.3 (vs. 81.0), MobileWorld 77.8 (vs. 51.2), Parametric CAD Bench 91.5 (vs. 73.8), RecreationBench 51.7 (vs. 30.2).
  • Document & Office Intelligence: CharX Intelligence): CharXiv (RQ) 88.4/93.5 (vs. 85.8/85.9), OmniDocBench 1.5 92.1 (vs. 91.4), OCR‑Bench‑V2 74.2/68.3 (vs. 70.7/67.1), CC‑OCR‑Bench‑V2 79.6 (vs. 72.7), MTVQA‑Test 56.6 (vs. 51.2), MADQA 91.8 (vs. 87.1).
  • Real‑World & Spatial Understanding: RealWorldQA 88.0 (vs. 86.9), ERQA 77.8 (vs. 69.8), LingoQA 84.8 (vs. 83.4), SURDS 77.8 (vs. 77.2).
  • Visual Perception & Grounding: SimpleVQA 75.0 (vs. 70.3), WorldVQA 53.2 (vs. 43.9), MMStar 85.9 (vs. 83.2), PerceptionBench 63.5 (vs. 51.1), CountQA 82.4 (vs. 77.0), RefAdv‑S 80.2 (vs. 73.0), Dense200 87.0 (vs. 60.7), COCO 78. 748.7 (vs. 74.2), VisFactor 60.8 (vs. 42.8), VLMsAreBiased 88.3 (vs. 36.6).
  • Video Intelligence & Agents: VideoMME (w/ Sub.) 90.4 (vs. 88.0), VideoMMMU 88.7 (vs. 85.4), MLVU (M‑Avg) 90.8 (vs. 87.4), TVBench 81.9 (vs. 78.2), LVBench 81.8 (vs. 76.2), LVBench (w/ Mem.) 85.6 (vs. 74.5), EgoLife (w/ Mem.) 80.3 (vs. 68.8), VideoDR (w/ Search) 73.2 (vs. 41.0).

These figures illustrate that Qwen3.8‑Max matches or exceeds leading models on many tasks while introducing open weights at the Max scale for the first time.

API Usage and Integration

Qwen3.8‑Max is accessible via QwenCloud. The API supports a reasoning_effort parameter to trade off depth, speed, and cost:

  • xhigh (default): thorough analysis for complex tasks.
  • medium: balanced accuracy and speed.
  • low: fast, cost‑effective reasoning. preserve_thinking is enabled by default for the best out‑of‑the‑box experience.

Example call using the OpenAI‑compatible client:

import os
from openai import OpenAI
api_key = os.environ.get("DASHSCOPE_API_KEY
if not api_key:
    raise ValueError("DASHSCOPE_API_KEY is required.
client = OpenAI(
    api_key=api_key,
    base_url=os.environ.get(
        "DASHSCOPE_BASE_URL",
        "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
    ),)
messages = [{"role": "user", "content": "Write a Python function to merge two sorted linked lists."}]
completion = client.chat.completions.create(
    model="qwen3.8-max",
    messages=messages,
    extra_body={
        "enable_thinking": True,
        # "preserve_thinking": True,
    },
    reasoning_effort="xhigh",
    stream=True,
)
# stream handling omitted for brevity

The model integrates with popular agent frameworks and coding assistants:

  • Claude Code: set ANTHROPIC_MODEL="qwen3.8-max", ANTHROPIC_BASE_URL=https://dashscope-intl.aliyuncs.com/apps/anthropic, and ANTHROPIC_AUTH_TOKEN=<your_api_key>.
  • Codex: add a model entry to ~/.codex/model-catalog.local.json with slug: "qwen3.8-max", context_window: 1000000, supported_reasoning_levels covering low, medium, xhigh, and point model_provider to ModelStudio with the appropriate base URL and API key.
  • Qoder CLI: install via curl -fsSL https://qoder.com/install | bash then run qoder.
  • Qwen Code: npm install -g @qwen-code/qwen-code@latest then qwen.
  • OpenClaw: install via curl -fsSL https://molt.bot/install.sh | bash, set DASHSCOPE_API_KEY, and configure ~/.openclaw/openclaw.json to modelstudio provider.modelstudiowith the QwenCloud endpoint and the model IDqwen3.8-max`.

Community Reaction

Comments on the Hacker News post highlight several points of interest and skepticism:

  • "They've also announced Qwen3.8-27B being released open-weight next week. Qwen3.6-27B is widely regarded as one of the best local models… If 3.8 truly improves upon it that would be awesome." – @toshinoriyagi
  • "3.8 27b is the real news here." – @boredatoms
  • "I don't understand… That's dated today, but… Qwen3.8 is launching and going open-weight soon! … You don't have to wait to test it. Just now, the Qwen3.8-Max-Preview made its debut on Alibaba’s Token Plan, Qoder, and QoderWork." – @simonw
  • "Can a model be stripped off anything not relevant to coding and get a lot lighter? Or is that impossible?" – @kopirgan
  • "I hope this is significantly cheaper. I've been loving Deepseek for its nearly free usage costs, hard to justify switching from cents per day." – @ddxv (referring to the reasoning_effort levels for cost control)
  • "Lmao I love their video with the idea that people will be able to do their hobbies while ai does their job. Surely Alibaba is leading by example here by reducing work hours per week while keeping pay the same right? Right?" – @jofzar
  • "I'm trying and failing to find value running a potential Qwen 3.8 27b dense model on a 16 core, 128 GB of ram, 2080ti box… How are you all justifying economical use of these local models right now?" – @aliljet
  • "Unfortunately the alibaba cloud does not seem like it can keep up with the demand caused by this announcement, i'm just getting endless timeouts." – @efficax

These community remarks reflect excitement about the forthcoming 27B open‑weight variant, curiosity about cost‑efficiency and model specialization, and practical concerns regarding cloud latency and local hardware feasibility.

Conclusion

Qwen3.8‑Max sets a new bar for autonomous AI systems by combining massive scale (2.4T parameters, 95B active) with open‑weight availability, strong self‑evolving coding loops, broad real‑world work proficiency, long‑horizon optimization in hardware and business simulations, and multimodal vision that acts as a continuous feedback loop. Its performance on coding challenges, research reproduction, competitive contests, professional workflows, chip design, e‑commerce benchmarks, and multimodal benchmarks demonstrates that the model can take open‑ended goals from start to finish with minimal human oversight and produce dependable, production‑ready artifacts. The upcoming release of its weights will enable the broader community to replicate and extend these capabilities.

Sources