Liquid AI released LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, bidirectional encoders built on the LFM2 hybrid architecture and available on Hugging Face. They support an 8,192-token context and are designed for fine-tuning on classification and token-level tasks. On CPU, LFM2.5-Encoder-230M is the fastest model tested from 1K tokens up, running about 3.7x faster than ModernBERT-base at 8,192 tokens.
Kimi K3 is now available on Nebius Token Factory, which is named a Day 0 launch partner, through an OpenAI-compatible API and console. The quoted post says Artificial Analysis scores the open-weight model at 57 on its Intelligence Index, two points behind GPT-5.6 Sol (max), and lists up to 1M tokens of context.
AIWhy it matters: The source names the cloud access route and an Artificial Analysis score of 57, letting readers compare Kimi K3 against GPT-5.6 Sol.
We've open-sourced MoonEP, our high-performance communication library for distributed MoE workloads. Built to make expert-parallel communication more efficient at scale, MoonEP helps reduce communication overhead in large MoE training and inference systems. Explore on GitHub: http://github.com/MoonshotAI/MoonEP
We've open-sourced AgentENV in collaboration with kvcache-ai. AgentENV is a distributed system for running agent environments at scale. Its components power agentic RL training for Kimi K3, with fast snapshot, resume, and fork support for large-scale parallel agent workflows. Explore on GitHub: http://github.com/kvcache-ai/AgentEnv
We've open-sourced FlashKDA, our high-performance CUTLASS-based implementation of Kimi Delta Attention kernels. It delivers 1.72×–2.22× prefill speedup over the flash-linear-attention baseline on H20, and works as a drop-in backend for flash-linear-attention. Explore on GitHub: http://github.com/MoonshotAI/FlashKDA
Moonshot AI is releasing the model weights and technical report for Kimi K3, a 2.8T-parameter MoE model with native visual understanding and a 1M-token context window. The post says the new architecture delivers 2.5x the intelligence per unit of compute, and the company is also opening high-performance attention kernels, an MoE communication library, and infrastructure for running agent environments at scale.
AIWhy it matters: The source names the model size, context window, and released weights, which helps readers compare its scale and openness with other frontier releases.
Kimi K3 (open weights, coming soon)
free the parameters 🤗
CohereLabs has published North-Mini-Code-1.0-eagle, an Eagle draft model for its North-Mini-Code-1.0 target model. The three-layer draft uses dense transformer blocks and sliding window attention, and it is compatible with fp8 and w4a16 target checkpoints in vLLM.
AI-based research and engineering is one of the main topics we're researching in the STAR Lab at Berkeley, and something I'm seeing used in industry more and more too. This is nice work from my grad students packaging multiple AI-based "autoresearch" algorithms into one API in the GEPA package so you can mix-and-match them to get the best results. These can be used for anything from writing prompts to designing agents to optimizing code.
BAAI has released AREX-Base, a 122B-total, 10B-activated Mixture-of-Experts deep research agent built on Qwen3.5-122B-A10B with a 262,144-token context. The model uses an inner research loop and an outer self-improvement loop, and the source reports it scoring 82.5 on BrowseComp and 85.4 on GAIA, under Apache 2.0.
AIWhy it matters: The release pairs a 122B-parameter deep research agent with benchmark tables against frontier and open models, letting readers compare its search-agent results directly.
BAAI's AREX-Turbo is a dense 4B deep research agent built on Qwen3.5-4B with a 262,144-token context length. It scores 70.7 on BrowseComp, 81.6 on GAIA and 40.6 on HLE with tools, versus 82.5, 85.4 and 52.4 for the 122B AREX-Base. The model is released under Apache License 2.0 and targets lower-cost research-agent deployment.
Andrew Ng and Rohit Prasad announced OpenWorker, an open-source agent that produces deliverables such as documents, Slack messages, and calendar updates across files and everyday tools. It checks in before consequential actions, runs on Mac with Windows support coming soon, and works with user-supplied API keys for models including GPT 5.6 Sol, Claude Fable, Gemini 3.6, open-weight models, or local Ollama models. Source code is available on GitHub, and the tool requires the user's own API key.
Poolside released Laguna S 2.1, an open-weight model with 118B total parameters and 8B active per token. The author says it performs strongly on agentic coding and long-horizon tasks, and it can run on a single NVIDIA DGX Spark. Weights are on Hugging Face under the OpenMDW-1.1 license, with access also available through OpenRouter and Poolside's API.
Create Artifacts directly from Claude Tag!
Introducing the UPGRADED Skywork Posters Create and publish posters that fit all social platforms in just one click.
Start your next video with Skywork Video 👉 https://skywork.ai/video
Run upgraded Gemma 4 with llama.cpp (QAT + MTP)
Mistral AI released Shieldstral-1.0-3B, a 3B-parameter multimodal safety classifier that judges content against natural-language policies and outputs a continuous safety score. It moderates text, image, and text-plus-image content in a single forward pass and can be retargeted to new policies at inference time without retraining. The Apache 2.0 open-weight model is built on Ministral-3-3B-Base-2512 and trained on sequences up to 32k tokens.
nice work and great to see new arch working! btw the small model is not opensourced?
Thinking Machines introduced Inkling, a model that reasons across text, image, and audio, and is making its full weights available. It is available today for fine-tuning on Tinker and can be tried in the Inkling Playground. John Schulman says pretraining began last winter and a small team added coding, reasoning, and agentic training starting in mid-January.
AIWhy it matters: The post links an open-weights release to a stated training timeline, showing how a small team moved from pretraining to coding, reasoning, and agentic training.
Thinking Machines introduced Inkling, an open-weight model with 975B parameters that reasons across text, image, and audio. The full weights are available, with fine-tuning on Tinker and access through the Inkling Playground and Hugging Face and partners.
Mira Murati announced Inkling, the first model from Thinking Machines, trained from scratch with its full weights made available. The model reasons across text, image, and audio, and is available today for fine-tuning on Tinker and testing in the Inkling Playground.
Inkling is the first in a family. We’ve included some details of Inkling-Small, a lighter-weight model trained on a similar recipe, with full weights to follow.
Liquid AI released Antidoom, an open-source method that retrains a single overtrained token to eliminate "doom loops" in small reasoning models. On LFM2.5-2.6B and Qwen3.5-4B, loop rates fell from 10.2% to 1.4% and from 22.9% to 1%, respectively. The company also released IFStruct, an open-source benchmark measuring whether model outputs satisfy a schema, where LFM2.5-350M rose from 21.10% to 44.90% after training.
What Stability AI audio researcher CJ Carr (@cortexelation) is covering in the workshop: 🎵 Run the model locally (even on a MacBook Pro) 🎵 Train your own LoRA 🎵 Use Stable Audio with Ableton 🎵 Build your own audio tools on top of the model … and more
If you are a producer, creative coder, or just curious about how this tech actually works under the hood, this is for you. Basic comfort with music tools and Python is helpful, but not needed.
the 10000th release of llama.cpp 🥳
ByteDance's UniVR-34B-Planning, built on Emu3.5 at 34B parameters, learns visual reasoning, physical dynamics, and long-term planning from visual demonstrations using a next-token objective and two-stage training on the VR-X dataset with VR-GRPO reinforcement learning. On the VR-X benchmark it scores 58.2 overall, up 18.4 points from the Emu3.5 34B baseline of 39.8. The Planning checkpoint is available on Hugging Face under CC BY 4.0, alongside a General checkpoint.
Checkout the latest version of Zed, adding proper support for llama.cpp as a model provider
Very excited for this model. It’s an enormous improvement over composer 2.5 and trained entirely from scratch. It’s been a pleasure working with the SpaceXAI team on it. Even more excited by the slope of the effort and the models to come.
llama.cpp recently added DFlash support to its speculative decoding arsenal. Along with MTP, Eagle3 and various ngram-based techniques, the local model performance takes another step up. Special thanks to NVIDIA team and Ruixiang Wang specifically for leading this effort! https://github.com/ggml-org/llama.cpp/pull/22105
Tufa Labs won the first $37.5K ARC-AGI-3 milestone prize with "The Duck," a small open-source LLM that plays the games by writing and running Python in a live REPL. Reki placed second with a vision-language agent using Gemma-4-31B, and md Boktiar Mahbub Murad placed third with the "forge" framework. The second and final milestone prize ends September 30.
A model repository named XiaomiMiMo/MiMo-V2.5-DFlash is listed on Hugging Face with 311B parameters and tensor types F32, BF16, and F8_E4M3. The README is empty, and the page reports 434 downloads last month and no Inference Provider deployment.
Most AI audio models have never heard a maqam. Team Motif fine-tuned Stable Audio 3.0 on Arabic maqam, built an Ableton plugin for microtonal style transfer, and won our Stable Audio 3.0 Challenge at Music Hackspace running locally on device. Watch Jad Al Masri break it down 👇
Bridgewater, one of the worlds largest hedge funds, a Tinker customer talks through how they've carefully fine-tuned a model focused on what makes interesting financial news. Their fine-tuned model is more effective and cheaper than any frontier model.
DeepSeek has posted the eagle3_gemma4_12b_ttt7 model on Hugging Face, listed at 2B parameters in BF16 tensor format with 621 downloads last month. The model card is empty and no Inference Provider currently deploys it, while it appears in the DeepSpec collection of 12 items.
DeepSeek published a model named eagle3_qwen3_14b_ttt7 on Hugging Face, listed at 2B parameters in BF16 with the Safetensors format. The model has no model card, is not deployed by any Inference Provider, and recorded 400 downloads last month.
DeepSeek has published eagle3_qwen3_8b_ttt7 on Hugging Face, a model listed at 2B parameters in BF16 safetensors format. The model card is missing, it has 577 downloads last month, and no Inference Provider currently deploys it.
DeepSeek has published eagle3_qwen3_4b_ttt7 on Hugging Face, a 0.9B-parameter model in BF16 tensor format. The model has no model card and is not deployed by any Inference Provider, and it is listed in the DeepSpec Collection.