On August 3, Alibaba released Qwen3.8-Max — the flagship model tuned for coding and agent work. The news, in a day, gathered nearly 600 points on Hacker News, and there are two reasons for it. First, the numbers: in several measurements, the model surpasses OpenAI and Anthropic’s closed flagship models. Second — and this is crucial for our audience — Alibaba for the first time promised to open the weights of the Max-class model: the giant with 2.4 trillion parameters and the compact Qwen3.8-27B are expected to go into open access within a week. If this happens, the bar for “frontier model = closed API” will fall completely.
Architecture and characteristics
Qwen3.8-Max is a sparse MoE-model (mixture-of-experts): 2.4 trillion parameters in total, ~95 billion active per token — activation rate about 4%. This is exactly the approach DeepSeek and Kimi use to achieve flagship quality without flagship inference costs. The rest by spec:
- context 1 million tokens, output up to 128k;
- multimodality (text + images) with a “native visual planning” — the agent can rely on screenshots;
- controllable reasoning levels (low / medium / high / xhigh) — you can trade quality for price and speed;
- API: $2 / $6 per million tokens (input/output) and $0.25 for cached — in agent sessions where 90% of the cost is rereading codebase, cheap cache matters more than the base price. For comparison: cached input is costlier for Claude and GPT by orders of magnitude.
Numbers and claims
From independent measurements that have already appeared: 87.3% on SWE-bench versus 82.6% for GPT-5.5, 4th place in Frontend Code Arena (1668 Elo), 2nd in Vision Arena, second place among open-weight models in the Vals index (66.1). Alibaba itself emphasizes long autonomous runs: a heartbeat that ran for ~16 days without human involvement during development with self-evolution of bindings; optimization of the chip layout from 8298 to 678 GBytes over 500+ iterations; an annual e-commerce simulation with a capital increase of 4.16x.
This should be taken seriously. Even in July, when the model lived in Max-Preview status, independent reviewers noted: “second place after Fable 5” — this is a positioning, not an reproducible result: neither the model specs nor the measurement configurations were published at that time. In the HN thread there are live examples of imperfections: one tester’s model drew an SVG bicycle in 11 minutes and $0.17, produced it without wheels, even though wheels were described in code. The old rule remains: other people’s benchmarks are a reason to test, not to migrate.
What practitioners say
The most valuable thing in the discussion is the experience of those who have long used Qwen locally. A few telling signals from the HN thread: the previous generation Qwen3.6-27B/35B is already the main local model for many; one developer wrote they canceled a Claude subscription as far back as spring; the 27B-dense version is often considered higher quality than 35B-MoE at smaller size; on a Mac Studio with 64 GB memory, a model of this class occupies ~20–25 GB. That is, the announced Qwen3.8-27B will arrive on a prepped ground: in 4-bit quantization it will fit on a single 24 GB GPU or a Mac with 32+ GB unified memory. The full Max is not about self-hosting: you need at least 8× H100/B200 for it.
There is also skepticism, and it is practical: DeepSeek users don’t see the point in paying $2/$6 where they are used to “almost free”; geographic restrictions in the license are discussed (USA, EU, UK, Korea) — before the final license text is published, embedding the model in a commercial product is risky.
What this means for us
If you run coding agents via API — add Qwen3.8-Max to tests already now (access via Alibaba Cloud Model Studio): a million-context and cheap caching make it a candidate specifically for long sessions. If your profile is local inference, wait for weights 27B and measure on your tasks: if it inherits even part of the agent capabilities of the senior model, this will be the most interesting local model for coding today. And in both cases — wait for the license before building anything seriously.
Sources
- Official announcement of Qwen3.8-Max
- HN discussion (~600 points): practitioner experience
- Analysis of the release in AINews / Latent Space
- NxCode: why to treat the announced Preview benchmarks with caution
Who tried Qwen (3.6 or 3.8) in coding agents — how does it compare with Claude/GPT on real tasks, not benchmarks? And are you waiting for 27B weights to finally switch to local inference?