Alibaba Qwen3.8 Open-Sources Flagship Model with 2.4T Parameters
Nashnova编辑部
Alibaba released Qwen3.8 as open weights for the first time at the flagship Max tier — 2.4T total parameters — breaking its long-standing practice of reserving top models behind an API. Developers can now deploy, modify, and commercialize the full-scale model directly.
Why is this open-source release different?
Every previous Qwen generation followed a two-track rule: smaller models (e.g. 72B) got open weights; the flagship Max was API-only.
Qwen3.8-2.4T is the first time Alibaba has released a Max-tier model as open weights.
This means → Alibaba is voluntarily dismantling the walled garden around its flagship. The competition shifts from selling API calls to winning the developer ecosystem.
What does a 2.4T-parameter model look like?
92 layers, 2.4T parameters — but only 95B activate per token. This is a MoE architecture — mixture of experts, splitting the model into 512 specialist sub-modules and routing to just 10 per inference, saving compute.
The attention mechanism is a "hybrid" design: 69 layers of linear attention + 23 layers of full attention, roughly a 3:1 ratio. In plain terms = most layers run on a fast, compute-light track; every fourth layer switches to a high-precision "slow lane" — balancing speed and quality.
Native context is 262K tokens, extensible to roughly 1 million tokens, with multi-token prediction (MTP — predicting 3 candidate tokens at once instead of one-by-one) training.
How similar is it to Moonshot's Kimi K3?
K3's 93 layers use the same split: 69 linear-attention layers + 24 full-attention layers, also roughly 3:1. The two architectures are strikingly convergent.
The key difference: K3 adds an attention-residual (AttnRes) mechanism that lets the model recombine outputs from multiple earlier layers. Qwen3.8 sticks with a more conventional Pre-Norm residual structure.
This reflects a broader pattern: flagship models in the 2–3T parameter class are converging fast on architecture, leaving less room for structural differentiation.
How far has the deployment bar dropped?
At full precision (BF16/FP8), the model requires at least two Nvidia B300 or AMD MI355X nodes.
Third-party framework Inferact offers FP4-quantized versions that compress it down to a single node. In plain terms = from two top-tier servers to one — still expensive, but the deployment bar is halved.
Inference combines tensor parallelism (TP — splitting matrix computation across GPUs), data parallelism (DP), and expert parallelism (EP — distributing experts across GPUs), with fused communication kernels to cut scheduling overhead.
Where did the post-training effort go?
Qwen lists Agent execution as its core upgrade — autonomous planning, environment-feedback handling, end-to-end task completion.
Benchmark gains concentrate on agent tasks: Terminal Bench 74.5 → 86.6, PaperBench 64.8 → 93.0, JobBench 31.3 → 53.4. Traditional single-turn knowledge reasoning (GPQA Diamond) barely moved, from 92.4 to 92.6.
This means → the competitive frontier for flagship models is shifting from "answer questions better" to "work independently." Marginal returns on single-turn QA are thinning out; long-horizon agents are the new battleground.
What's hidden in the open-source license?
The open-weights version is text-only, thinking-mode only, 262K context. The cloud-hosted Max adds vision input, non-thinking mode, default 1M context, and built-in tools — it is not a full replica.
The license changed from Apache 2.0 to a bespoke Qwen3.8-Max License: commercial use is allowed, but products with over 100 million MAU or over $20 million monthly revenue must prominently display the model name. MaaS or AI-assistant businesses with over $50 million in annual revenue need a separate authorization.
Kimi K3 adopted a similarly customized license with comparable thresholds. This reflects an emerging consensus among top open-source models: small teams use it freely; large companies pay or attribute — the definition of "open source" is being redrawn.
Content is for reference only, not financial advice.