--- title: Qwen3.8-Max: what we know, what's unverified, and what... description: Alibaba has launched Qwen3.8-Max, a 2.4-trillion-parameter sparse mixture-of-experts model activating ~95B parameters per inference, with a 1-million-token... canonical: https://daily.dev/posts/qwen3-8-max-what-we-know-what-s-unverified-and-what-developers-are-actually-saying-tce0j5xea twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Qwen3.8-Max: what we know, what's unverified, and what developers are actually saying | daily.dev og:description: Alibaba has launched Qwen3.8-Max, a 2.4-trillion-parameter sparse mixture-of-experts model activating ~95B parameters per inference, with a 1-million-token... og:url: https://daily.dev/posts/qwen3-8-max-what-we-know-what-s-unverified-and-what-developers-are-actually-saying-tce0j5xea og:image: https://api.daily.dev/og/posts/tCe0J5XEa.png og:image:alt: Qwen3.8-Max: what we know, what's unverified, and what developers are actually saying og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Qwen3.8-Max: what we know, what's unverified, and what developers are actually saying **[Collections](https://daily.dev/sources/collections)** · 4 min read · 4 upvotes · 0 comments ## Summary Alibaba has launched Qwen3.8-Max, a 2.4-trillion-parameter sparse mixture-of-experts model activating ~95B parameters per inference, with a 1-million-token context window and multimodal support. Self-published benchmarks place it near Claude Opus 4.8 and GPT-5 on coding and agentic tasks, but independent verification is pending. A headline demo showed the model autonomously building a CLI tool over 16 days with a public GitHub history, though developers note the results depend heavily on controlled scaffolding rather than raw capability. Open weights for the full model and a 27B variant are promised on Hugging Face but haven't shipped yet, drawing criticism that it's currently an API product with an open-source label. Enterprise considerations around data sovereignty, indemnification, and the practical limits of long context windows in production are also flagged. ## Content Alibaba launched Qwen3.8-Max this week, and the AI community has had a lot to say about it — some impressed, some skeptical, most somewhere in between. ## What it is Qwen3.8-Max is a sparse mixture-of-experts model with 2.4 trillion total parameters, though only about 95 billion activate per token. That distinction matters: you get the stored knowledge of a 2.4T model while paying for roughly 95B worth of compute per inference. It supports text, images, and video, with a 1-million-token context window and a maximum output length of 131,072 tokens. The internal "thinking budget" before the model commits to an answer stretches to 262,000 tokens. Pricing on QwenCloud is $2 per million input tokens and $6 per million output tokens, with cached reads dropping to $0.17 per million — about 8x cheaper if you're reusing a stable prompt prefix. That's significantly cheaper than comparable closed-lab models, though at least one developer noted Kimi K3 runs agentic tasks at $0.86 per run versus Qwen3.8-Max's $1.14, because the Alibaba model takes more turns and pushes output token usage up roughly 45%. Open weights for both the full 2.4T model and a smaller 27B variant are expected on Hugging Face and ModelScope next week. The model is already available via QwenCloud, OpenRouter, and Venice. ## Benchmarks On Terminal Bench 2.1 (which tests whether a model can actually drive a real command line through a task end to end), Qwen3.8-Max scored 86.6, compared to 84.6 for Claude Opus 4.8 and 88.8 for GPT-5.6 Sol. It posted 93.0 on PaperBench and 92.6 on GPQA Diamond. On Arena's Image-to-WebDev leaderboard, it entered at #2, behind Claude Opus 5 but ahead of Fable 5. On the open-weights SWE-bench leaderboard, Kimi K3 still leads at 57 points, with Qwen3.8-Max one point behind. The honest summary: it's competitive with top-tier models on several benchmarks, but not clearly ahead of any of them. Most analysts treating Kimi K3 and Qwen3.8-Max as roughly tied on capability, with cost and deployment context being the real differentiators. ## The 16-day coding demo Alibaba's headline demo had the model autonomously build a CLI tool called "oh-my-cli" over 16 days, producing 265 commits, 127 pull requests, and 151 GitHub issues — all publicly auditable on GitHub. Other demos include reproducing a research paper's results in five days (about 7,600 lines of code, 33 training rounds, all 6 findings reproduced) and optimizing a cryptographic circuit from 8,298 logic gates down to 678 over roughly 500 iterations, cutting chip area by 81% while still meeting timing at 500 MHz. Those are genuinely interesting results. But developers have pointed out something worth sitting with: Alibaba didn't just hand the model a task and walk away. It built a tightly controlled environment — specific tools, constrained interfaces, structured feedback loops — and the model operated inside that cage. The insight isn't that the model is autonomous; it's that the scaffolding did a lot of work. Capacity is not the same as judgment. ## What developers actually think Reactions have been mixed in ways that feel pretty honest. On the open-weights promise: the weights haven't shipped yet, and until they do, this is an API product. One developer called it "an API business model wearing an open source jacket." That's a bit harsh, but the point stands — promised weights and available weights are different things. On the benchmarks: Alibaba published same-day results from its own evaluations. Several developers flagged this as "a lab grading its own homework." Independent verification is still pending. On the million-token context window: useful for prototyping, but developers with production experience caution that long-context performance degrades quietly. Stuffing a million tokens into a prompt is not a production architecture. On pricing: some developers find it competitive; others are sticking with GPT-5.6 Sol plus DeepSeek Flash as a cheaper combination. The 27B variant, when weights drop, may be more practically deployable for most organizations than the full 2.4T model, which is not realistically self-hostable for most teams. On the broader picture: open-model benchmark scores have gone from around 2% on Humanity's Last Exam a year ago to over 50% now. Whatever the caveats around Qwen3.8-Max specifically, the trajectory of open-weight models is real. ## The geopolitical read Alibaba released this model despite ongoing US export controls on advanced chips. Whether Qwen3.8-Max represents a genuine frontier capability or a well-optimized near-frontier model is still being debated, but the gap with Western labs has clearly narrowed. That's the part of this story that will matter longest. ## Questions this post answers ### What are the pricing details for Qwen3.8-Max on QwenCloud? Qwen3.8-Max costs $2 per million input tokens and $6 per million output tokens on QwenCloud, with cached reads dropping to $0.17 per million tokens, about 8x cheaper for reused prompt prefixes. Despite this, one developer found Kimi K3 cheaper for agentic tasks, at $0.86 per run versus $1.14, because Qwen3.8-Max takes more turns and uses roughly 45% more output tokens. _Comparing model pricing across real workloads, not just list prices, is easier when tracking developer benchmarks on daily.dev._ ### How does Qwen3.8-Max compare to Claude Opus and GPT-5.6 Sol on Terminal Bench 2.1? Qwen3.8-Max scored 86.6 on Terminal Bench 2.1, versus 84.6 for Claude Opus 4.8 and 88.8 for GPT-5.6 Sol. It also posted 93.0 on PaperBench and 92.6 on GPQA Diamond, and trails Kimi K3 by one point on the open-weights SWE-bench leaderboard, where Kimi K3 leads at 57. Most analysts treat Qwen3.8-Max and Kimi K3 as roughly tied on raw capability. _Weighing benchmark claims against real coding tasks matters when picking a model, something daily.dev readers track closely._ ### Are the open weights for Qwen3.8-Max available yet? No, open weights for the full 2.4 trillion parameter model and a smaller 27B variant are only expected on Hugging Face and ModelScope the following week after launch; they have not shipped. Until they do, Qwen3.8-Max is accessible only as an API product through QwenCloud, OpenRouter, and Venice, prompting one developer to call it 'an API business model wearing an open source jacket.' _Tracking when promised open weights actually ship helps developers plan self-hosting decisions around models like this._ ## Similar posts on daily.dev - [Qwen 3.8 Released : Better than Kimi K3?](https://daily.dev/posts/qwen-3-8-released-better-than-kimi-k3--gsoqcloby) · Medium · 1 upvotes · 0 comments --- Tags: [#open-source](https://daily.dev/tags/open-source), [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#qwen](https://daily.dev/tags/qwen), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts) [View this post on daily.dev](https://daily.dev/posts/qwen3-8-max-what-we-know-what-s-unverified-and-what-developers-are-actually-saying-tce0j5xea) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Qwen3.8-Max: what we know, what's unverified, and what developers are actually saying","url":"https://daily.dev/posts/qwen3-8-max-what-we-know-what-s-unverified-and-what-developers-are-actually-saying-tce0j5xea","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/qwen3-8-max-what-we-know-what-s-unverified-and-what-developers-are-actually-saying-tce0j5xea"},"datePublished":"2026-08-04T14:50:42.852Z","dateModified":"2026-09-13T19:43:09.051Z","description":"Alibaba has launched Qwen3.8-Max, a 2.4-trillion-parameter sparse mixture-of-experts model activating ~95B parameters per inference, with a 1-million-token...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/706e61521e6cc1d77e4395be7e98cc26?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/706e61521e6cc1d77e4395be7e98cc26?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/qwen3-8-max-what-we-know-what-s-unverified-and-what-developers-are-actually-saying-tce0j5xea","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":4},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"open-source,llm,ai-agents,qwen,mixture-of-experts","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Qwen3.8-Max: what we know, what's unverified, and what developers are actually saying"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/qwen3-8-max-what-we-know-what-s-unverified-and-what-developers-are-actually-saying-tce0j5xea#faq","mainEntity":[{"@type":"Question","name":"What are the pricing details for Qwen3.8-Max on QwenCloud?","acceptedAnswer":{"@type":"Answer","text":"Qwen3.8-Max costs $2 per million input tokens and $6 per million output tokens on QwenCloud, with cached reads dropping to $0.17 per million tokens, about 8x cheaper for reused prompt prefixes. Despite this, one developer found Kimi K3 cheaper for agentic tasks, at $0.86 per run versus $1.14, because Qwen3.8-Max takes more turns and uses roughly 45% more output tokens. Comparing model pricing across real workloads, not just list prices, is easier when tracking developer benchmarks on daily.dev."}},{"@type":"Question","name":"How does Qwen3.8-Max compare to Claude Opus and GPT-5.6 Sol on Terminal Bench 2.1?","acceptedAnswer":{"@type":"Answer","text":"Qwen3.8-Max scored 86.6 on Terminal Bench 2.1, versus 84.6 for Claude Opus 4.8 and 88.8 for GPT-5.6 Sol. It also posted 93.0 on PaperBench and 92.6 on GPQA Diamond, and trails Kimi K3 by one point on the open-weights SWE-bench leaderboard, where Kimi K3 leads at 57. Most analysts treat Qwen3.8-Max and Kimi K3 as roughly tied on raw capability. Weighing benchmark claims against real coding tasks matters when picking a model, something daily.dev readers track closely."}},{"@type":"Question","name":"Are the open weights for Qwen3.8-Max available yet?","acceptedAnswer":{"@type":"Answer","text":"No, open weights for the full 2.4 trillion parameter model and a smaller 27B variant are only expected on Hugging Face and ModelScope the following week after launch; they have not shipped. Until they do, Qwen3.8-Max is accessible only as an API product through QwenCloud, OpenRouter, and Venice, prompting one developer to call it 'an API business model wearing an open source jacket.' Tracking when promised open weights actually ship helps developers plan self-hosting decisions around models like this."}}]} ```