Thinking Machines ships Inkling 2 Mini: a 26B looped MoE with Apache weights, built for 32GB machines
Inkling 2 Mini runs 3.4B active of 26.4B params, loops its middle layers three times, takes 1M tokens and fits a ~20GB checkpoint. Apache 2.0.
Topic · 87 stories · page 1 of 4
87 stories tagged open-weights on AI Tamer, newest first.
Inkling 2 Mini runs 3.4B active of 26.4B params, loops its middle layers three times, takes 1M tokens and fits a ~20GB checkpoint. Apache 2.0.
The two open-weight image models to know: MiniMax H3-Image for 16 MP and 16 references, HiDream-O1 for an MIT licence at 8B. Ranks, hardware and licences.
Prime Intellect's Model Factory beta runs RL, SFT, pre-training and one-click endpoints for open-weight models, per GPU-hour or on dedicated clusters.
Hugging Face Inference Providers now send LLM calls via OpenRouter, with retries and :fastest or :cheapest routing. Your own provider keys no longer work.
Prices rechecked 14 Nov. Hume's old TTS docs now redirect, Chatterbox's README calls the repo an archive, and the main API prices still match.
LiquidCodec turns 24 kHz speech into 12 tokens per 80 ms frame and back, loads in Transformers, and uses Liquid's revenue-capped LFM Open License.
LTX-2.5 is open weights with per-second API prices up to 4K. Vidu Q4 Preview takes 15 reference images and ranks higher on image-to-video. Checked 12 Nov.
PAN 2 pairs a language model that plans the next world state with a 29B video model that renders it, up to 20 seconds. Weights and code are Apache 2.0.
Praxis-1 weights, inference and fine-tuning code are public under Apache 2.0. The quick start needs one 80 GB GPU; Runway's benchmark numbers are its own.
MiniMax H3-Image makes and edits images up to 20 megapixels with up to 16 references. The API is live and the weights are out under a restrictive licence.
ZONOS3-Mini is an 800M-parameter MoE text-to-speech model with about 100M active, Apache-2.0 weights, voice cloning and runtimes for browsers, Macs and SGLang.
HyperFlow H3 Flash is Video Rebirth's few-step MiniMax-H3 for joint video and audio. The H3 community license still sets the territory limits.
On 5 November ASD's ACSC published two papers on open-weight models: inventory them, check hashes and signatures, and approve risky actions outside the model.
Nous fine-tuned Qwen3.8-27B on 52.8M tokens and says propaganda answers fell from 22% to 0.9%. Apache-2.0 weights, eval and data are public.
Linum's P-JiT 21K is a 2.18B pixel-space image model trained on 8,415 H100-hours. Apache-2.0 weights and code; it cannot render text.
Three open-weight TTS picks: Kokoro's tiny Apache model, Resemble's MIT Chatterbox family with cloning, and Fish Audio Drama 3, the top open model on AA.
Underdog's Woof TTS: ~31.7 MB / ~30.7M params, WebGPU in-browser demo at underdog.ai/escape. License CC BY-NC 4.0. No confirmed Hub GA.
Inkling 2 Small: 594B MoE / 17B active with vision, on Tinker. AA Index about 31. Apache-2.0 weights now public on Hugging Face after safety work.
NVIDIA posts four Cosmos 3 Nano open-weight checkpoints for interactive driving and robot world simulation under OpenMDW1.1.
Viggle's Remake adds two LoRAs to MiniMax H3 to swap people, objects, weather or style in existing video. The weights carry H3's licence, which bars US use.
Hindsight 1.2 can hand small memory judgments to hs-decide-1-flash, an Apache 2.0 0.8B model. Vectorize says some operations get 2.5x cheaper; it is opt-in.
Baidu released a 280B-total, 14B-active pre-training checkpoint on Hugging Face. Its custom license bars use in the EU, UK and South Korea.
Two open-weight video models: MiniMax H3 adds stereo sound, 2K and up to 12 references; HunyuanVideo 1.5 runs on a 14 GB GPU. Prices, ranks, licences.
Ricursive put Ripple-65B Preview on Hugging Face under Apache 2.0. Cognition's SWE-2 Flash, built on an open base, withholds them. Scores are self-reported.