benchmarked Qwable-5-27b-coder on my rtx 5090. a real-trace coding SFT of Qwen3.6-27b, trained on Claude Fable-5 traces then Kimi 2.7 Coder traces. every cheap eval agrees it's fine: ~ quality q_avg…

I ran qwen3-tts-1.7b vs fish-s2-pro - 150 seed-tts-eval english utterances. these models are TTS: you feed text + a short reference clip of a voice, and they generate that text spoken in that voice…

new model drop: Qwable-v1, a Qwen3.6-35B-A3B distilled on Claude Opus-4.7 reasoning, then SFT'd on Claude Fable-5 agentic traces. I ran it controlled against its own vanilla base on one 5090, same…

ran all three speculative-decoding drafters for Gemma 4 MTP vs EAGLE-3 vs DFlash model used is 26B-A4B single stream, median of 3, over a 193 tok/s baseline: DFlash 2.19x · MTP 2.13x · EAGLE-3 1.69x…

my four verdicts on Qwopus3.6-27B-Coder: (1) quality real. q_avg 94.1 across MMLU/ARC/HellaSwag/GSM8K/HumanEval - ranks #2 on my thinking-off leaderboard, above its own base family (94.0) at a…

I wanted to have a look at hour-video models, so I checked Kuaishou's Keye-VL-2.0-30B. it is apache 2.0, "lossless 256K context", beats Qwen3-VL-235B on LongVideoBench. I spent the full day trying to…

my synthetic agentic leaderboard was too thin. so I tightened it: all 7 models, 30 harder SWE-bench Verified bugs, real tools in docker, graded by the official swebench harness. r drops to 0.50…

I benched Nex-N2-mini (the agentic post-train of Qwen3.5-35B-A3B) against its own base: native tool-calling, 15 deterministic tasks scored by a programmatic checker. I built an "agentic score" so the…

Mistral small 4 is a 119 billion parameter model. it runs on my 32GB VRAM GPU, at ~36 tokens/sec of actually-usable generation. I compared it with an other model in its lane. both on one RTX 5090,…

I tried to do FP8 reinforcement learning on my 5090 and got a real +7.66 on GSM8K! ~~~ the initial plan was the fast one: unsloth's FP8 RL (load_in_fp8=True), vLLM rollouts, ~1.4x faster. on…