Thinking Machines' Inkling 2 Small Trains a 594B Model in 68 Hours
Thinking Machines open-sourced a 594B parameter Mixture-of-Experts model pretrained end-to-end in 4-bit NVFP4 on Blackwell GB300 racks in 68 hours.
20 stories
Thinking Machines open-sourced a 594B parameter Mixture-of-Experts model pretrained end-to-end in 4-bit NVFP4 on Blackwell GB300 racks in 68 hours.
Cactus Compute's Mirage packs scene description, detection, OCR, segmentation and structured extraction into a 22 MB model that runs on-device in 74 ms.
MiniMax released a 33B parameter open-weights image model that generates and edits 16 megapixel images in just 6 sampling steps.
AutoTrust's 27B decision model outputs calibrated option probabilities instead of free text, topping tool selection and intent classification benchmarks.
Hindsight AI released a 114M-parameter multilingual cross-encoder reranker that runs transformer compute at 15M parameters and handles 8K-token pairs.
Alibaba's smallest Qwen3.5 base model packs native vision, 262K context, and a Gated DeltaNet hybrid stack into under a billion parameters.
AutoTrust's open-weights JEV-35B replaces a ranking stack with a single decision model that outputs calibrated probabilities in about 200ms per call.
Comfy-Org has repackaged MiniMax's open-weight H3-Image model into quantized single-file checkpoints that drop straight into ComfyUI workflows.
Meta's new promptable model reconstructs a full-body 3D human mesh including hands and feet from a single photo, with keypoint and mask prompting baked in.
Edge0 ships a 145B latent Mixture-of-Experts model derived from Qwen3.8-Flash-Next that runs on a single 24GB consumer GPU at nearly 100 tokens per second.
NVIDIA ships a Blackwell-ready NVFP4 build of Moonshot's 2.8T parameter multimodal MoE, with accuracy effectively matching the full-precision reference.
A quantized GGUF build of AdaVaR-3B lets you run a Qwen2.5-VL fine-tune that picks between text reasoning and image-grounded reasoning on the fly.
A 27B coding model distilled on 195M tokens of agentic debugging traces lands in GGUF form, runnable on a single workstation GPU.
A compact German named entity recognizer built on XLM-RoBERTa and document-level context hits 92.31 F1 on CoNLL-03 and crosses 948K downloads.
A quantized GLM-5.3 variant trims 64 experts via neural architecture search, shrinking the model 24.6% so it fits on a single B200 GPU.
An ONNX port of BAAI's 33M-parameter embedding model runs entirely in the browser through Transformers.js, hitting 850K monthly downloads.
UkisAI released mixed-precision GGUF quantizations of Swift Flash Next that cut thinking tokens by 63% while keeping accuracy loss under 1%.
NVIDIA's 16B omnimodal world model unifies text, image, video, audio, and robot action generation under one Mixture-of-Transformers architecture.
Moondream's Parakeet Ultra speech-to-text model now ships as GGUF quantizations for transcribe.cpp, delivering 1.8% WER at 486 MB across 25 languages.
Zyphra's new open-source TTS model runs real-time voice cloning on phones and laptops with 100M active parameters and Apache 2.0 weights.