Pruna AI Squeezes a 143 GB Video Model Into 21 GB
Pruna AI converted MiniMax-H3 video model into an open image generator that runs on 21 GB of VRAM at 5x the speed.
20 stories
Pruna AI converted MiniMax-H3 video model into an open image generator that runs on 21 GB of VRAM at 5x the speed.
A self-hosted TypeScript middleware that lets Claude Code, Codex and other CLI agents run autonomously in the background and ping you only for key decisions.
A 40-plus author collaboration shows how to tune vision-language models for Southeast Asia while keeping global performance intact.
169Pi's 9B vision-language model runs offline on laptops and phones, hitting 61.20% on SWE-Bench Verified with Apache 2.0 weights.
A single illustration plus a coding agent becomes a face-tracked VTuber avatar that blinks, talks, and streams straight into OBS.
OrcaRouter shipped GGUF builds of an abliterated 40B MoE coding model that run on a 16GB or 32GB GPU with 262K context.
A new supervision recipe forces models to actually use retrieved evidence, not just paste it next to a prediction, with large gains in radiology verification.
A community retrofit grafts Qwen Sparse Attention onto Qwen3.8-27B, nearly tripling prefill speed at 256k context while matching dense quality.
A community repack of Alibaba's 8-step image generator ships in GGUF, FP8, NVFP4 and MLX formats, hitting 2.5M downloads and 4.5k likes.
Google's open medical vision-language family built on Gemma 3 matches specialist models while staying cheap enough to fine-tune on hospital data.
A new personality benchmark finds frontier closed models write with more truthfulness and structure, while open weights lean into filler and bottom-line summaries.
A local, open-source skill gives Claude Code, Cursor and Codex 42 FFmpeg tools with a machine-readable contract, no cloud, no API keys.
A new arXiv paper tackles cell lineage reconstruction when only a fraction of cells are labeled, combining 2.5D and 3D detectors with joint lineage optimization.
A partially decensored, vision-capable GGUF build of Swift 1.5 Qwen3.8-27B that still scores 97.6% on GSM8K and runs locally in 25 GB VRAM.
Artificial Analysis is opening up the CritPt v2 physics benchmark grading server via its API, alongside a small Intelligence Index revision.
Anthropic's FlashProve generates cryptographic proofs that a response came from a specific LLM, running 70x faster than prior open-source systems.
Phone Harness ships an open-source Python toolkit that gives coding agents real taps, swipes, and typing on iPhones and Androids.
Prime Intellect audited popular LLM benchmarks and found broken graders, ambiguous tasks, and web-based cheating that significantly change reported model rankings.
Topaz Labs ships Wonder 4, a mixture-of-experts image model that upscales, denoises, sharpens, deblurs, and recovers faces in a single pass.
Jane Street open-sourced a Socratic tutor skill that drills LLMs into question-driven teachers with diagnostics, level maps, and spaced review.