Qwen 3.8 Max Explained: Alibaba's 2.4 Trillion Parameter Model
Qwen 3.8 Max is Alibaba's open-weight 2.4 trillion parameter model with frontier coding and agentic benchmarks. Here's what it can actually do.
What is Qwen 3.8 Max?
Qwen 3.8 Max is Alibaba’s newest and largest AI model, a 2.4 trillion parameter system with 95 billion active parameters, built to be multimodal from the ground up. It’s the biggest release yet in the Qwen family, and Alibaba has said it will open-source the weights about a week after the initial announcement, making it the largest open-weight model available once that happens. It’s positioned as a direct competitor to closed frontier models like Anthropic’s Fable line and OpenAI’s GPT 5.6 Soul, particularly on coding and long-running agentic tasks.
TL;DR
- Qwen 3.8 Max is a 2.4 trillion parameter model with 95 billion active parameters, making it the largest model Alibaba has released and set to become the largest open-weight model once weights go public.
- On agentic coding benchmarks like Terminal Bench and SWE-Bench Pro, it performs competitively with Fable and GPT 5.6 Soul, though it trails on some harder, more realistic tests like Deep SWE 1.1.
- Alibaba highlighted a demo where the model ran autonomously for over 16 days, building its own coding harness (OMI CLI) from an empty repository, complete with its own state machine, dispatcher, and testing loop.
- In a simulated year-long e-commerce benchmark, the model quadrupled its starting capital, managing multiple online stores, negotiating with suppliers, and spotting scam vendors in the process.
- The model is Anthropic API compatible, meaning it can be dropped into tools like Claude Code or Codex as a substitute backend.
- Pricing on its first-party API undercuts Western frontier models significantly, though token efficiency (not just per-token price) determines real-world cost.
- A smaller 27 billion parameter version is also coming, aimed at developers who want to run a capable model on consumer hardware.
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
How big is Qwen 3.8 Max, really?
At 2.4 trillion total parameters, Qwen 3.8 Max sits in the same size class as Moonshot AI’s Kimi K3, which also launched around 3 trillion parameters shortly before it. Only 95 billion of those parameters are active at inference time, which is standard for large mixture-of-experts style models: most of the network stays dormant for any given token, keeping compute costs down relative to a dense model of the same size.
For context, closed frontier models from US labs are rumored to run significantly larger. Fable is estimated at 7+ trillion parameters, roughly double Qwen’s size, and OpenAI’s next major training run is expected to land in similar territory. That gap matters: Chinese labs are producing models that are highly competitive on benchmarks while using a fraction of the parameters and, by extension, a fraction of the compute and chip access that US labs can draw on.
What benchmarks does Qwen 3.8 Max lead or lag on?
The picture is mixed, which is normal for a model this size, but the pattern is consistent across sources.
On agentic coding benchmarks, Qwen 3.8 Max lands close to Fable and ahead of GPT 5.6 Soul on Terminal Bench 2.1 and SWE-Bench Pro, putting it in the top three among leading models. It’s weaker on Deep SWE 1.1, a benchmark considered more realistic and harder to game. It reportedly beats both Fable and GPT 5.6 Soul on PaperBench, a test that measures how well a model can reproduce the results of a research paper from scratch, no starter code, no pipeline, just the paper and a set of GPUs.
On the LM Arena leaderboard, it ranks fourth overall, making it the second open-weight model to break into the top five. Breaking benchmarks down by category shows Qwen leading in multimodal reasoning, document and office intelligence, real-world and spatial understanding, and visual perception, while Fable still leads on visual agent and coding-specific tasks.
The consistent theme: Qwen 3.8 Max is elite at long-horizon, instruction-following, grunt-work style agentic tasks. On the hardest, most demanding engineering benchmarks, it’s strong but still trails the top Western models. Worth noting: most of these benchmark numbers come from Qwen’s own agentic harness, not independent third-party testing, so real-world performance under different tooling may vary.
What can it actually do in long-running tasks?
The headline demo is a project Alibaba calls OMI CLI. The model was given a completely empty code repository and, without human hand-holding, worked on it autonomously for roughly 16 days, according to Alibaba’s own account. Rather than just writing code, it appears to have built something resembling its own engineering organization: a state machine, a dispatcher, a monitor, and a watchdog operating as one loop. Requirements land as issues, the agent claims them, moves them through stages like ready, in-progress, and active, then triggers end-to-end tests and CI checks before merging pull requests.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
Separately, Alibaba ran a simulated e-commerce benchmark based on Taobao and Tmall data. The model started with 100,000 in capital and had to run multiple online stores in parallel for a full simulated year, negotiating with suppliers in natural language, adjusting prices, handling returns, and navigating disruptions like typhoons. Out of a pool of suppliers, 152 were scammers the model had to identify. It finished with a balance of roughly 416,000, quadrupling its starting capital and reportedly beating a comparable open model (GLM 5.2) by 38%.
The model was also used for an autonomous chip design task: given a working but bloated design using 8,298 logic gates, it iterated roughly 500 times and reduced the design down to 678 gates, an 81% reduction in area.
These long-horizon tasks matter more than single-question benchmarks because errors compound. A model that makes a small mistake early in a 365-day simulation or a 16-day autonomous build will usually see that mistake snowball. Qwen 3.8 Max completing these tasks with improving results over hundreds of iterations is a meaningfully different capability signal than acing a one-shot coding question.
Is Qwen 3.8 Max worth using over closed models?
It depends heavily on the use case. On pricing, Qwen’s first-party API reportedly runs around $2 per million input tokens and $6 per million output tokens, compared to roughly $5/$30 for GPT 5.6 Soul and $10/$50 for Fable. That’s a substantial discount on paper.
But per-token price isn’t the whole story. If a model needs three or four times as many tokens to complete the same task, the effective cost converges with pricier models. Independent cost-per-completed-task tracking (from firms like Artificial Analysis) hasn’t yet incorporated Qwen 3.8 Max, so a clean apples-to-apples cost comparison isn’t available yet.
The model is also Anthropic API compatible, meaning it can be plugged into existing agentic tooling built for Claude, such as Claude Code or Codex-style harnesses, as a drop-in backend. That compatibility lowers the switching cost for developers already building on those ecosystems.
A smaller 27 billion parameter variant is also planned, aimed at people who want to run a capable Qwen model on consumer-grade hardware rather than the trillion-parameter flagship, which would require enterprise-scale infrastructure to serve.
What are the risks of an open-weight model this capable?
Open-weight release means anyone who downloads the model has it permanently. Unlike a closed API, it can’t be pulled offline or restricted after the fact if a safety issue turns up. That’s the trade-off with releasing frontier-class weights: enormous benefit for researchers, startups, and tinkerers who get access to near-frontier capability for free, alongside the loss of any kill switch if the model turns out to enable serious misuse.
There’s also a geopolitical dimension. If US companies increasingly rely on cheaper, highly capable open models originating from Chinese labs, and those models are eventually optimized alongside Chinese chip architectures, dependency risk grows even when inference is run entirely on US soil. That’s a separate question from whether the model itself is safe to use, but it’s part of the broader conversation around Qwen 3.8 Max and similar releases like Kimi K3.
Frequently Asked Questions
Is Qwen 3.8 Max open source?
It’s open-weight, meaning Alibaba plans to release the trained model weights publicly (roughly a week after the initial announcement) so anyone can download and run them, though the exact licensing terms, including any commercial-use restrictions, weren’t confirmed at launch.
How does Qwen 3.8 Max compare to Fable and GPT 5.6 Soul?
One coffee. One working app.
You bring the idea. Remy manages the project.
It’s competitive on several agentic coding benchmarks like Terminal Bench and SWE-Bench Pro, and leads on some multimodal and document-understanding tasks, but it still trails on harder, more realistic coding benchmarks like Deep SWE 1.1.
Can I run Qwen 3.8 Max on my own computer?
The full 2.4 trillion parameter model requires large-scale infrastructure to run, not consumer hardware. A smaller 27 billion parameter version is planned specifically for developers who want something runnable on local or consumer-grade GPUs.
What is OMI CLI?
OMI CLI is the autonomous coding harness Qwen 3.8 Max built and evolved over roughly 16 days of self-directed operation, starting from an empty code repository, and which Alibaba has also released publicly.
Why does Alibaba highlight chip design and research paper reproduction?
Both are seen as precursor capabilities to more advanced automated research and hardware self-improvement. A model that can accurately reproduce a research paper’s results from scratch, or iteratively optimize a chip layout, demonstrates the kind of step-by-step, self-correcting reasoning needed for more autonomous scientific and engineering work.
