KAT-Coder V2.5 (35B, NVFP4) on one DGX Spark: 64.6 tok/s decode. The fastest we have…
This is a dev post classified by AGK Intelligence as AI dev tools (a tool drop), kept by the Dev Radar because it carries real work, not commentary.
A new model for me! KAT-Coder V2.5 (35B, NVFP4) on one DGX Spark: 64.6 tok/s decode. The fastest we have measured on a single Spark. 21.9 GB of NVFP4 weights with a grafted MTP head, served with vLLM 0.24.0 at 262,144-token context. Recipe and build by sakamakismile (GitHub). 4 slots, exactly as the recipe specifies. Decode code: 64.6 tok/s Decode prose: 64.6 tok/s 500 words: ≈ 11.4 s C1: 62.4 tok/s C2: 105.5 tok/s C4: 131.6 tok/s C8: 130.1 tok/s MMLU-Pro: 57.3% ± 4.8 GPQA: 37.9% ± 6.7 IFEval: 77.5% ± 5.8 LiveCodeBench: 35.0% ± 14.2 Ruler 8K: 80.0% ± 12.1 Ruler 32K: 62.5% ± 14.4 Setup took
Posted by Niklas (2.2k followers) 3 days ago · 4 likes · 1.1k views · view the original post on X. Kept by the Dev Radar as AI dev tools.
More dev work like this
- 🎃Dario给中国人民带来的极致恐怖从Claude无情封号开始 — @NFT_Chen
- Mia's Qwen3.8-Flash-Next recipe on a DGX Spark: one command, first try, 2 min to the… — @niklaslenz_ai
- 给AI喂一整仓，最烦的就是它每回都从头翻文件，这毛病你多半也遇到过。 — @bkdgiffug
- Coding agents aren't slow because they're dumb. They're slow because every edit waits on… — @RunAnywhereAI
- ICYMI: Foursquare co-founder @dens sat down with our Head of Product, AI, Jeff Smith at… — @automattic
- 🤯惊了！9B参数大模型，只需16G内存即可本地跑通，能力反超Qwen3.5-27B！ — @NFT_Chen
- Introducing JetSpec++: we find causal parallel drafting paired with hidden states… — @haoailab
- Build the agent, test the agent, grade the results, change the agent, run the test cases… — @rseroter
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 31.2k posts from 5k X accounts over the last 21 days, 8.8k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-10-30 20:27 UTC. Full method.