Kimi K3.1 Pro vs Qwen3.8 2.4T-A95B

Analysis

Kimi K3.1 Pro and Qwen3.8 2.4T-A95B are the only two open-weight models past two trillion parameters, and they need almost exactly the same hardware: one node of eight 288 GB B300s. That makes this the one comparison where hardware cannot decide for you. What can is context length, input types, licensing, and how much you trust each vendor's own benchmarks, because the two have never been run side by side on the same harness.

Both are mixtures of experts with hybrid attention. Kimi has 2.8T parameters with 104B active, Qwen 2.4T with 95B active, and both keep a per-token cache on only about a quarter of their layers: 24 of 93 for Kimi, 23 of 92 for Qwen, with linear-attention layers holding a fixed-size state in the rest. At an 8k context Kimi's native MXFP4 checkpoint needs 1,749.1 GB and Qwen's Q4_K_M estimate needs about 1,669 GB. There is one difference in how solid those numbers are. Kimi's figure is the file Moonshot actually ships (1,561.3 GB of weights); Qwen's is this site's estimate for a community GGUF, since Alibaba's own quantized release is FP8. On eight B300s this site estimates the same 24.2 tokens per second for both at 8k. Eight 180 GB B200s hold neither. Context is where they split. Qwen's native window is 262,144 tokens (Alibaba documents settings that stretch it to about 1M); Kimi's native window is 1,048,576. At Qwen's 262,144 tokens, Kimi's KV cache is actually the smaller of the two, 12.9 GB against 24.7 GB, because each Qwen cache entry is twice as wide, so the totals are 1,763.1 GB against 1,696.2 GB. Inputs split too: Kimi takes text and images, Qwen is text only. On benchmarks, the only suite both report is GPQA-Diamond, 94.2 for Kimi against 92.6 for Qwen, each from its own vendor. Qwen also reports Terminal-Bench 2.1 (86.6) and SWE-bench Pro (67.7); Moonshot reports neither for K3.1 Pro and leads with Terminal-Bench 4.0 (65.7), a harder, different suite, so no coding comparison can be drawn. Licenses are close: Qwen's allows commercial use, asks for attribution above $20M a month in revenue or 100M monthly users, and needs a separate license only for model-as-a-service businesses over $50M a year; Kimi's requires a separate agreement for model-as-a-service businesses past $20M over 12 months.

Bottom line: If you have the B300 node, the hardware is a wash, so choose on the workload. Kimi K3.1 Pro is the pick for image input, for native million-token context, and for anyone who values Moonshot's reported drop in hallucinations. Qwen3.8 2.4T is the pick if your agents already run in Claude Code style harnesses, where Alibaba measured its coding scores, or if a looser revenue threshold on model-as-a-service use matters to you. Until someone runs both on the same benchmarks, treat any claim that one is clearly smarter as unproven.