Alibaba's Qwen2.5-3B Shrinks to 2.5GB so Anyone Can Run AI Locally
A 4-bit quantized build of Qwen2.5-3B-Instruct runs chat, coding and long-context tasks on roughly 2.5 GB of VRAM.
20 stories
A 4-bit quantized build of Qwen2.5-3B-Instruct runs chat, coding and long-context tasks on roughly 2.5 GB of VRAM.
A community fine-tune of Qwen3.8-27B trades a sliver of raw accuracy for shorter reasoning traces, faster decoding, and cleaner Python output in long agent loops.
Recursive's first model is a 40B mixture-of-experts coder that activates just 2.5B parameters, runs on a MacBook, and was largely designed by an automated research agent.
Thinking Machines Lab's 276B-parameter MoE matches its 975B flagship across reasoning, coding, and multimodal tasks while using 12B active parameters per token.
Lodestones ships Kromoss, a full fine-tune of OpenMOSS's 5B diffusion transformer at 1024px, bundled with 8-step turbo LoRAs that beat the base model on PartiPrompts.
A community LoRA strips the guardrails off OpenMOSS's 5B text-to-image model, turning a layout-focused DiT into an adult-content generator under Apache 2.0.
Hugging Face's Carbon-X teaches a 14B Qwen3 model to speak DNA, matching GPT-6.1 Sol on genomics benchmarks at 1/1,800 the cost.
Z.ai's 106B Mixture-of-Experts model activates just 12B parameters per token, trades a sliver of depth for cheaper inference, and ships under MIT.
A community-fine-tuned Slovak speech recognition model built on Meta's XLS-R has quietly racked up 860k downloads on Hugging Face.
A 125B mixture-of-experts coding model squeezed onto a single 24GB consumer GPU through evolutionary expert pruning and 2.125-bit trellis quantization.
Tencent Hunyuan has released Hy-Nomas, an open interactive video world model that generates 720p footage one segment at a time under live text and camera control.
Taobao's Content AI team squeezes MiniMax-H3 reference-to-video-plus-audio generation into 4 or 8 denoising steps using per-hop LoRA adapters.
An 8-step distilled MiniMax H3 that turns image, video, and voice references into clips with synced audio now runs natively in ComfyUI.
A community port brings VLX-Seek's open-vocabulary detection VLM to pure C++ with GGUF quantization, hitting 2.2x decode speedup and running in 12GB VRAM.
A 27B fine-tune of Qwen3.8 built for everyday messages beats much larger models at sounding like an actual human being.
Base Labs post-trained Qwen3.5 into two open-weight models that output calibrated probabilities over discrete choices instead of generated text.
Baseten's Base Labs open-sources a 9B decision model that returns calibrated probabilities instead of text, topping its weight class on the Jev Decision Index.
SparkLLM's new 1.7B on-device model nearly matches its 4B sibling on coding and agent benchmarks, with native 1M context and controllable thinking.
Liquid AI's new 24B mixture-of-experts activates only 2B parameters per token, delivers 184 tokens/sec on an M5 Max, and ships in GGUF for llama.cpp.
Bria's 8B Fibo 2.0 unifies image generation and editing, tops open-weights leaderboards, and ships trained only on licensed data.