Liquid AI's LFM2.5 Hits 600 Tok/s but Fails Every Quality Check
A community experiment converts Liquid AI's 24B mixture-of-experts model into a block masked-diffusion sampler, hitting 600 tokens/second on one A100 but with broken outputs.
20 stories
A community experiment converts Liquid AI's 24B mixture-of-experts model into a block masked-diffusion sampler, hitting 600 tokens/second on one A100 but with broken outputs.
An open in-context LoRA on Hugging Face reconstructs MiniMax-H3 drafts at 2K or 4K in three to four steps, filling a gap MiniMax left closed.
Continuity Research released Tess-4.5-27B, a Qwen3.8-27B fine-tune that matches its base on BenchLocal using 25% fewer tokens and runs locally.
A community distill pours Claude Opus and Sonnet reasoning traces into Liquid AI's 1.7B tri-hybrid model, packaged as GGUF quants that fit in about 1GB.
Concept Country rewrote nine weight matrices in Qwen3.8-27B with no training, cutting reasoning tokens up to 54% while raising benchmark scores.
IBM's third-generation TinyTimeMixer brings mixture-of-experts routing, trend-residual decomposition, and roughly 15 to 50 times faster forecasting than popular SOTA baselines.
Ricursive Intelligence trained a 65B dense coding model for $32M that beats open-weight MoE giants up to 15x its size on agentic coding.
A community GGUF port brings Tencent's DSpine draft model to llama.cpp, pushing Qwen3.8-27B decoding to ~105 tok/s on a single RTX 3090.
Nous Research took Alibaba's Qwen3.8-27B, post-trained it on 53M tokens, and cut state-aligned political framing from 22% to under 1%.
AutoTrust Labs released a 27B text decision model that scores option probabilities instead of generating free-form answers, hitting 97 on tool selection.
A 122B mixture-of-experts reasoning model with MIT weights, a 262K context, and 84.5% on SWE-bench Verified, served from a single 80GB GPU.
Prism ML's 2B vision-language model squeezes into roughly half a gigabyte using ternary weights while keeping a 262K context window and image input.
A solo developer fully fine-tuned Krea 2's 12.8B base model on 196k anime illustrations in about 8 hours, then transferred the result to the fast 8-step Turbo checkpoint.
A specialized 4B Qwen fine-tune writes complete Three.js pages that preview low-poly assets and export them as GLB files ready for Godot.
A telecom-specialized 31B model post-trained from Gemma 4 on 440B tokens of standards data lands on Hugging Face with QLoRA efficiency.
Qwen3-ASR-1.7B now ships as GGUF quantizations for transcribe.cpp, bringing a state-of-the-art multilingual speech recognizer to local CPU inference.
A hobbyist cloned Wispr Flow into a free, local Mac app by fine-tuning LiquidAI's LFM3-1.7B with DeepSeek-generated targets.
MIT Han Lab's EfficientViT-L2 swaps softmax attention for multi-scale linear attention, cutting cost on high-resolution vision while staying accurate.
A Vietnamese sentence embedding model fusing SimCSE contrastive learning with PhoBERT has quietly crossed 850K downloads on Hugging Face.
Prism ML squeezes a 9B vision-language model down to 2 GB with ternary weights, hybrid attention, and 262K context for on-device inference.