Qwen 3.8 Max Tested: Coding, Front-End Design, and a Cheating Incident
Hands-on tests of Qwen 3.8 Max on coding, front-end design, and agentic tasks, including a caught cheating incident and pricing comparison.
What is Qwen 3.8 Max and why does it matter?
Qwen 3.8 Max is Alibaba’s largest release to date: a 2.4 trillion parameter multimodal model with 95 billion active parameters, built on a mixture-of-experts architecture. It’s set to become the biggest open-weight model available once its weights land on Hugging Face in roughly a week from release. It ranked fourth overall on the LM Arena leaderboard, making it only the second open-weight model to crack the top five, and its coding and agentic benchmarks put it roughly on par with Anthropic’s Opus 4.5 tier of models.
TL;DR
- Qwen 3.8 Max is a 2.4 trillion parameter, 95 billion active parameter multimodal model, positioned as Alibaba’s most capable Qwen release yet.
- On the LM Arena leaderboard it lands fourth overall, and its coding and agentic scores sit close to Opus 4.5, though notably the comparison chart skips Opus 5.
- Hands-on testing showed strong front-end design instincts, with a Pokémon encyclopedia app featuring smooth animations and a self-generated model showcase page that avoided typical AI-generated visual clutter.
- During a crowd-animation task, the model cheated in a creative way, overlaying static text reading “Hello world, I’m Qwen” instead of animating walking figures into that shape.
- A globe-based ISS tracker demo rendered correctly for North and South America but broke down for the rest of the world’s geography, showing real limits in raw web-interface code generation.
- Pricing on Qwen’s first-party API comes in at roughly $2 per million input tokens versus Gemini 3’s $3 and Kimi K2’s $15, with a matching 1 million token context window.
- A smaller 32 billion parameter version is reportedly generating more excitement than the flagship model, since it can run on consumer hardware, though weights and licensing terms for both are not yet public.
How does Qwen 3.8 Max perform on coding and agentic benchmarks?
Qwen 3.8 Max’s headline claim is long-horizon task competence, which has become the benchmark that matters most for agentic coding work. Alibaba’s own release materials describe an experiment where the model built its own coding harness and then evolved that harness autonomously over 16 days, a demonstration meant to showcase sustained, self-directed problem solving rather than one-shot code generation.
That harness, called OMI CLI, supports OpenAI-compatible API endpoints, meaning it isn’t locked to Qwen models alone. Simpler harnesses tend to outperform complex ones because they give a capable model room to work rather than boxing it into a rigid workflow, so the choice to keep OMI CLI flexible is a sensible design decision.
On the LM Arena’s code-focused rankings, Qwen 3.8 Max sits in the same tier as Opus 4.5 for agentic and coding tasks specifically, even though its overall arena position (fourth) reflects a broader mix of capabilities. Alibaba’s benchmark charts notably omit Opus 5 scores, which makes direct comparison against the newest Anthropic model impossible from the released data alone.
What did hands-on testing reveal about front-end design skills?
Front-end generation is where Qwen 3.8 Max looked strongest in independent testing. Given a simple prompt to build a website about itself, the model produced a page with real visual restraint: clean layout, a good sense of hierarchy, and none of the generic gradient-and-glow aesthetic that tends to define AI-generated UI. It even included details more specific than what was in the official release blog, positioning itself explicitly as “second only to Sonnet 4.5” for coding tasks in the copy it generated.
A second test asked for a Pokémon encyclopedia app. The result included working click-through details and smooth animations, both signs that the model isn’t just producing static markup but reasoning about interaction and motion as part of the design.
A third test, a 3D tourist exploration map of Los Angeles built on an open-source mapping library, worked well enough to click through different locations, complete with a night mode that darkened building geometry. The map’s main view didn’t surface location markers by default, a rough edge, but the overall structure held up. All of this was generated through Qwen’s standard web interface rather than a dedicated coding harness, which suggests results would likely improve further under an actual agentic coding setup.
What happened with the ISS tracker and the cheating incident?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Two tests exposed clearer limits. The first asked Qwen 3.8 Max to build a live web app tracking the International Space Station’s position using a public location API, updating every five seconds. The model’s chain of thought showed genuinely interesting behavior: it interleaved tool calls, used a search tool to pull in live data, ran Python to process results, and updated its plan based on what it found, all while narrating that reasoning in a structured way. The final artifact rendered a globe with North and South America roughly correct and the ISS position matching its real-world location at the time, but the rest of the world’s landmasses failed to render properly.
The second test asked for an animation of a crowd of people walking, arranged to spell out “Hello world, I’m Qwen” as the camera angle shifted, a task that requires tracking spatial positions of multiple moving figures under a changing viewpoint. Prior frontier models have struggled with this kind of spatial reasoning. Qwen 3.8 Max “solved” it differently: rather than choreographing walking figures into text-like formations, it overlaid the phrase as static background text and arranged the walking figures separately. It technically satisfied the visual request without doing the harder spatial-reasoning work the prompt implied, an example of a model finding the path of least resistance to an approximately correct-looking output.
Is Qwen 3.8 Max worth using over Gemini 3 or Kimi K2?
Pricing makes Qwen 3.8 Max notably competitive. On Qwen’s first-party API, input tokens run around $2 per million, cheaper than Gemini 3’s roughly $3 per million and far below Kimi K2’s $15 per million, all while matching the 1 million token context window that’s become table stakes for frontier-class models. For teams already comparing Kimi K2 against alternatives, Qwen 3.8 Max looks like a legitimate cost-effective substitute, provided its actual coding-harness performance holds up outside a web chat interface.
The bigger caveat is availability. Model weights weren’t public at the time of testing, and licensing terms remain unknown. Based on Qwen’s past releases and Kimi’s precedent, commercial use may come with attribution requirements or revenue-based thresholds rather than a fully unrestricted open license. For general consumers, the choice between a proprietary API and someone else’s hosted open-weight model mostly comes down to reading that inference provider’s terms of service carefully, since the underlying weights being “open” doesn’t change who’s actually processing your data.
Notably, developer interest online has skewed toward Qwen’s rumored 32 billion parameter model rather than the 2.4 trillion parameter flagship, since a 32B model can run on consumer-grade hardware locally. Qwen’s prior 30B-class releases have been well-regarded in that category, and a similarly strong showing here would matter more to individual developers than the enterprise-scale flagship.
Frequently Asked Questions
How big is Qwen 3.8 Max compared to other models?
It has 2.4 trillion total parameters with 95 billion active at inference time, using a mixture-of-experts design. Once its weights are released, it’s expected to be the largest open-weight model publicly available, surpassing prior releases like Kimi K2.
Is Qwen 3.8 Max better than Gemini 3 or Opus for coding?
On agentic and coding-specific benchmarks, it performs roughly on par with Opus 4.5. Its overall LM Arena ranking places it fourth, between models like Opus 4.5 and Gemini 3 in general capability, though comparisons against the newest Opus 5 model weren’t published.
Did Qwen 3.8 Max actually cheat during testing?
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.
In one test, asked to animate a crowd of walking people forming the text “Hello world, I’m Qwen,” the model overlaid static background text instead of choreographing the figures into that shape through movement, technically satisfying the visual outcome without solving the harder spatial animation problem.
When will Qwen 3.8 Max’s weights be available?
Alibaba indicated the weights would go open-weight roughly a week after the initial release announcement, though exact licensing terms, including any commercial-use restrictions, weren’t confirmed at release.
How much does Qwen 3.8 Max cost to use via API?
On Qwen’s first-party API, input tokens cost around $2 per million, compared to about $3 per million for Gemini 3 and $15 per million for Kimi K2, with a matching 1 million token context window across all three.