My first upload is at 9,494 downloads. Every rung of my Qwen3.6 TTQ ladder sits below the base repo's best file at its size, 1.26× to 3.51× lower KLD. 3.93 bpw: GPQA 84.8 (official 86.0), MMLU-Redux 92.6 (93.3). Both within noise. huggingface.co/TokenTinkerer…

don't wait to buy more RAM and VRAM , it's only gonna get more expensive in the next few months. get anything with an attractive price on the current used market , old server/workstation ddr4 , nvidia/radeon/intel gpus , don't hesitate

Seven more packs are up BFCL v3, API-Bank, ARC-Challenge, ARC-Easy, TruthfulQA-817, AIME-2026 and HMMT-Feb-2026. Want a benchmark added? Drop it in the replies Install by URL, no Docker, run on your own model: github.com/TokenTinkererAI/b… @BenchLocalAI @stevibe

finished building the @BenchLocalAI packs i was talking about the other day. 12 out of the 13 are also available as prebuilt archive under releases. @stevibe github.com/TokenTinkererAI/b…

the whale pet is mandatory if you are using deepseek harness, i don't make the rules

Been using @BenchLocalAI by @stevibe, it's a great tool. Only issue: no 3rd party packs for deeper reasoning or coding benchmarks. So I'm porting well-known ones myself, starting with the ones that need no extra setup or verifiers. Will post when done

MiniCPM5-2B is impressive for it's size , in my tests it outperforms every 9b-12b models i have tested. it's not good at coding but no small model is. where it shines is how good at tool calling and maths it is for it's size , it fits on any hardware and it is fast

But results are interesting, nothing changes for small models , still runs on your fastest gpu. and for bigger models you can actually benefit from the extra vram : - Qwen 3.6 35B A3B Q4 K M - 5 tok/s ➡️ 25 tok/s without touching --n-cpu-moe -Qwen 3.6 27B - 3 tok/s ➡️ 17 tok/s

MoE broke the old "bigger = slower" assumption. A 35B MoE with 3B active params isn't as smart as a dense model of similar size , but it's much faster while still being smarter than the smaller dense models.