Alibaba just open-sourced its first Max-scale model
On August 13, Alibaba released the weights for Qwen 3.8-2.4T — the first time the Qwen team has open-sourced a model at Max scale. For years the strategy was a two-track setup: open-weight mid-size models, closed Max models served only through API. Qwen 3.8-2.4T ends that split.
The release itself is lean — no companion tech report yet, just architecture, post-training notes, inference infra, and a license. But read it next to Kimi K3 from two weeks ago, and the more interesting story appears: the open-weight flagship recipe has converged.
The recipe: hybrid attention at 3:1, ~100B active
Qwen 3.8-2.4T packs 2.4T total parameters with 95B active per token, spread across 512 experts (10 routed, one shared). The 92 layers are arranged as 23 repeated groups of three Gated DeltaNet layers plus one Gated Attention layer — 69 linear-attention layers and 23 full-attention layers, almost exactly a 3:1 hybrid. Kimi K3 runs the same 3:1 pattern with its 69 KDA layers and 24 Gated MLA layers. Same family, minor differences: KDA is a finer-grained take on Gated DeltaNet, and K3 adds attention residuals that Qwen 3.8 does not carry.
Native context is 262,144 tokens, extendable past 1M, with multi-token prediction training (three candidate tokens per step in the reference config) — the same MTP trick DeepSeek V4 documents in its report (see our DeepSeek V4 Pro API guide).
Inference infra is where the two diverge — and where the cost story lives
Official checkpoints come in BF16 and FP8; Inferact adds NVFP4 and MXFP4 quantized versions. Full-precision serving needs two B300 or MI355X nodes; FP4 fits on a single node. That is the number to watch: a 2.4T-parameter flagship running on one node changes who can host it.
Serving follows the now-standard split: tensor parallelism for the attention path, expert parallelism for the MoE path, with fused kernels and rack-scale interconnects keeping expert routing cheap. Kimi K3 attacked the same problem with MoonEP, dynamically replicating hot experts across GPUs based on load. The parallelism tuning, not the attention ratio, is the remaining differentiation surface.
Post-training: the entire delta went to agents
The model card is explicit: agent execution is the headline upgrade — autonomous planning, environment feedback handling, end-to-end task completion. Official numbers back it up: Terminal Bench 74.5 to 86.6, PaperBench 64.8 to 93.0, JobBench 31.3 to 53.4, Toolathlon Verified 49.7 to 72.5. GPQA Diamond moved only 92.4 to 92.6. Generation-on-generation gains have almost entirely shifted from knowledge benchmarks to agent benchmarks.
Just as telling: preserve_thinking is on by default, keeping reasoning context alive across tool calls. Kimi K3 made the same call. Reasoning state is moving from the harness into the model itself — a boundary shift worth watching closely (we covered the harness side of this debate in our MiniMax agent strategy breakdown).
Open-weight no longer means Apache
The license changed too. Qwen 3.8 ships under the Qwen3.8-Max License, not Apache 2.0: free to use, modify, fine-tune, and even sell derivatives, but products above 100M MAU or $20M monthly revenue must display the model name prominently, and MaaS or AI work-assistant businesses with $50M+ annual group revenue need a separate commercial license. Kimi K3's custom license follows the same shape. At the 3T class, open-weight plus custom license has become the de-facto standard.
Why this matters
Put Qwen 3.8 and Kimi K3 side by side and the 2026 flagship recipe is basically fixed: 2–3T total parameters, ~100B active, 3:1 hybrid linear attention, ~1M context, agent-first post-training, FP4/EP inference, open weights with tiered licensing. The remaining differences are micro-optimizations, not architectural bets.
For builders the practical takeaways: evaluate flagships on agent benchmarks, not GPQA; treat FP4 single-node serving as the new cost baseline for big models; and if you are building MaaS or an AI work-assistant product, read the license thresholds before integrating. The interesting question now is who breaks the converged recipe first.
This piece focuses on the recipe and inference side; the companion Qwen 3.8-2.4T Open Weights: Frontier AI Shift reads the same launch through benchmark results and industry structure. Read them together for the complete picture.
