Qwen3.8-Max: 2.4T Parameters Available Now, Open Weights Next Week

A model that writes code for 16 days straight. A model that designs a cryptographic chip from an empty folder, driving 8,298 gates down to 678 through 500 turns of self-directed optimization. A model that enters a live competition against 526 human teams and finishes ahead of 458 of them.

That model shipped today. Qwen3.8-Max is the Alibaba Qwen team's new flagship, and the open weights arrive next week.

The timing matters beyond the benchmarks. The open-weight model landscape just got a new ceiling, and for teams running inference on dedicated GPU infrastructure, the question shifts from "which model scores highest" to "what does it take to serve a 2.4T-parameter MoE at production throughput."

A 2.4T MoE with Open Weights

Qwen3.8-Max uses a mixture-of-experts architecture: 2.4 trillion total parameters with 95 billion active per token. On the Artificial Analysis Intelligence Index, Qwen3.8-Max scores 53, trailing Claude Opus 5 (61), Claude Fable 5 (60), GPT-5.6 Sol (59), Kimi K3 (57), and Grok 4.5 (54), but the open-weight release changes the deployment equation entirely.

Most frontier models at this tier are API-only. Qwen3.8-Max launches with API access via QwenCloud today and open weights next week. For teams that self-host or run on dedicated GPU infrastructure, that distinction is everything. You get the model weights, the architecture, and the freedom to serve it however your workload demands.

The model builds on the Qwen 3.5 architecture foundation and scales RL training across a combinatorial space of tasks, workspaces, and harnesses. The Qwen team describes three infrastructure innovations that made this scaling possible: continuously expanding real environments along independent axes, a universal reward system that unifies heterogeneous verification under automatically scalable rubrics, and an online data balancer that suppresses inter-batch gradient variance to sustain stable RL scaling.

The result is a model that performs at the frontier across coding, work, research, and multimodal tasks while maintaining consistent behavior across different agent harnesses including Claude Code, Codex, OpenClaw, and Hermes.

Coding: 16 Days, 265 Commits, Zero Human Intervention

The coding numbers are the headline, but the methodology is what makes them interesting.

On Terminal Bench 2.1, Qwen3.8-Max scores 86.6, placing it ahead of Claude Opus 4.8 (84.6) and behind GPT-5.6 Sol (88.8). On SWE-bench Pro, it hits 67.7. On FrontierSWE, 73.5. These are competitive with the best API-only models available.

But the Qwen team went further and ran three long-horizon autonomous coding experiments that stress-test what "coding capability" actually means at production scale.

Experiment one: The self-evolving harness. Qwen3.8-Max was asked to build oh-my-cli from scratch. Over 16 days of fully autonomous operation, it accumulated 265 commits, 127 PRs, and 151 issues. It built a complete engineering loop: requirements enter GitHub Issues, an agent claims them through a state machine, implements the code, triggers CI and E2E tests, and merges after passing. When tests fail, it routes failures back to the relevant issue for rework. It converted community feedback and developer reports into executable work items. The complete trace is public at github.com/qwen-code-dev-bot/oh-my-cli.

Experiment two: Reproduce a research paper, then beat it. Qwen3.8-Max was handed a recent paper on data selection for LLM reasoning, a set of GPUs, and the instruction to go. With zero starter code and zero pre-built pipeline, it wrote 7,600 lines of code, ran 33 rounds of GPU training, and reproduced all six main findings from the paper over 125 hours of continuous effort. It confirmed the paper's method beats random selection by +7.7% on AIME24. Then it kept going. Over four rounds of self-directed research, it invented and tested 18 improvement ideas. The best one, counting "hard decision points" in training examples, beat the paper's own method by +2.71 points.

Experiment three: Beat humans in a live competition. Qwen3.8-Max entered the WWW2025 Multimodal Dialogue Intent Recognition Challenge on the Tianchi platform, competing against 526 human teams. With a 24-hour time limit, it read the rules, built a solution combining fine-tuned BERT, RoBERTa, and MacBERT models with Qwen2.5-VL-7B for vision, constructed a weighted-voting ensemble, and iterated across 45 submissions. Final accuracy: 0.853, ahead of 458 teams.

These experiments test something that benchmark scores alone miss: sustained coherent decision-making over hundreds of turns without drifting or plateauing.

Designing a Chip, Gate by Gate

The chip design experiment is the most revealing test of long-horizon reasoning in the Qwen3.8-Max release.

The task: design a GCD/RSA cryptographic hardware accelerator in RTL. The model starts with empty module templates and a testbench. It has access to Iverilog for simulation, Yosys for synthesis, and OpenROAD for physical layout. It must maintain bit-exact functional correctness across 4-, 6-, 8-, and 16-bit configurations while minimizing gate count.

Over 500 turns and 71 evaluations across 13 key milestones, Qwen3.8-Max drove the design from 8,298 gates down to 678 gates. The optimization path shows genuine architectural reasoning rather than incremental tweaking:

- Turn 22: Algorithmic rewrite. It replaced an expensive 16-bit hardware modulo divider with an iterative shift-subtract architecture, slashing 6,288 gates in one move, over 80% of the total area reduction. - Turns 35-48: Redundancy elimination. It recognized that the caller's pre-conditions made a reduction stage unnecessary, merged two independent modules into one shared block, and narrowed internal register bitwidths. - Turns 60-113: Control logic pruning. It removed redundant registers, introduced an early-exit mechanism for even-number handling, and repurposed the subtractor's MSB as a comparator. - Turns 252 onward: Cross-module optimization. It inlined the multiplier directly into the modular exponentiation FSM, merged three sub-modules, and shared a single subtractor globally. - Turns 443-500: Gate-level refinement. Shared NOR-gate trees, absolute-difference subtraction splitting, and byte-to-bit selection logic squeezed out the final redundancies.

The physical layout tells the same story. The starting design occupied a 106x106 µm² die with 33,369 µm of wirelength and severe timing violations. The final layout shrank to 46x46 µm² with 4,187 µm of wirelength, achieving timing closure at 500 MHz. An 81% reduction in physical die area, driven entirely by the model's own architectural decisions, validated end-to-end through automated toolchains.

This is the capability that separates frontier models from the rest: the ability to make major structural breakthroughs hundreds of turns into a run, rather than plateauing after early low-hanging gains.

Multimodal Agents That Verify Their Own Work

Qwen3.8-Max brings vision into the agent loop as a native feedback mechanism. The model observes its own intermediate results during execution, inspects page layouts and spatial relationships, and corrects outputs when it detects misalignment between intent and result. Vision becomes a feedback loop across planning, execution, verification, and iteration, rather than just another input modality.

On multimodal reasoning benchmarks, Qwen3.8-Max scores 82.3 on MMMU-Pro, 91.9 on LogicVista, and 90.0 on HiPhO. On visual agent benchmarks, it hits 86.1 on OSWorld-Verified and 85.3 on AndroidWorld. These are competitive with Claude Fable 5 and ahead of GPT-5.6 Sol on several metrics. The full benchmark table, published in the Qwen3.8-Max release post, covers 50+ evaluations across text, vision, video, and agent tasks.

For document-heavy workflows, it processes financial reports and PDFs exceeding 200 pages, extracting insights across text, charts, and layouts. For video, it builds memory graphs from content longer than 100 hours, reconstructing event progressions and character relationships.

The Qwen team also introduces Qwen-MM-Plugins, a harness extension library that adds image and video processing, multimodal memory, and visual tool use to existing agent frameworks. Any agent harness can become multimodal-native through the plugin system. The library supports dynamic-resolution processing, fine-grained video memory, and specialized capabilities for video editing, Blender, and CAD workflows.

A new benchmark, RecreationBench, tests hybrid agent capability: the model observes a running application purely through interaction and feedback, with zero source code access, then rebuilds it from scratch through iterative coding and interactive verification. Qwen3.8-Max achieves frontier results across five platforms: Ubuntu, macOS, Windows, Android, and web.

What Open Weights Mean for GPU Infrastructure

The open-weight release next week is the signal that matters for infrastructure teams.

A 2.4T-parameter MoE model with 95B active parameters is a substantial inference workload. The full model requires significant GPU memory to hold all expert weights, but the active parameter count per token means that with expert parallelism and efficient batching, the per-request compute footprint is manageable on current-generation hardware. Running models at this scale on dedicated bare-metal GPU clusters gives teams full control over the serving architecture.

The Qwen team recommends reasoning_effort as the primary cost-control lever, with three levels: xhigh for complex tasks, medium for balancing accuracy and speed, and low for efficiency. Teams deploying on dedicated GPU infrastructure can profile each level against their workload mix and allocate hardware accordingly. A coding agent running at xhigh might use 8x H200 GPUs with full model parallelism, while a document processing pipeline at low could run on a fraction of that footprint with expert offloading.

A 2.4T-parameter model with open weights means teams own the serving architecture. Profile each reasoning level against your workload, route by complexity, and allocate GPU capacity where it actually matters.

Model routing becomes a natural complement. Low-complexity queries route through reasoning_effort=low on commodity GPUs. High-complexity coding and research tasks route through xhigh on high-memory instances. The open weights make this routing architecture possible without per-token API markup at every tier. For teams already running open-weight models on GMI Cloud, adding Qwen3.8-Max to the routing matrix means frontier capability alongside existing deployments.

The model also supports industry-standard protocols compatible with OpenAI and Anthropic APIs, integrating directly with Claude Code, Codex, Qoder CLI, Qwen Code, and OpenClaw. Infrastructure teams that have standardized on a particular agent harness can adopt Qwen3.8-Max without changing their toolchain. The Qwen team ships ready-to-use configuration snippets for each harness, reducing integration time to minutes.

For teams that have been running open-weight models and waiting for a frontier-capable option at this scale, next week is the starting gun. Qwen3.8-Max is available today via API through QwenCloud, and the weights land next week on Hugging Face and ModelScope. The GMI Cloud blog will track the open-weight release and inference benchmarks as they become available.

Run Qwen3.8-Max Today on GMI Cloud

Qwen3.8-Max is live on GMI Cloud's serverless inference API today, no waitlist, no dedicated cluster setup required. Point your existing OpenAI-compatible client at https://api.gmi-serving.com/v1/chat/completions, swap in the model ID, and you're running frontier inference in minutes:

curl --request POST \
  --url https://api.gmi-serving.com/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer YOUR_GMI_API_KEY' \
  --data '{
    "model": "Qwen/Qwen3.8-Max",
    "messages": [
      {"role": "system", "content": "You are a helpful AI assistant"},
      {"role": "user", "content": "List 3 countries and their capitals."}
    ],
    "temperature": 0,
    "max_completion_tokens": 500
  }'

Start Building Today

Join the GMI Cloud builder community on Discord or find us on X at @gmi_cloud.