Open Source AI

Moonshot AI Releases Kimi K2.5: An Open Source Visual Agentic Intelligence Model with Native Swarm Execution Kimi K2.5 is an open source visual agentic model from Moonshot AI that targets coding, multimodal reasoning, and research automation. It uses a Mixture of Experts architecture with 1T total parameters, about 32B active parameters per token, 61 layers, 384 experts, and a 256K context length. A MoonViT vision encoder with about 400M parameters and training on about 15T mixed vision and text tokens give it strong document and image understanding. Agent Swarm, trained with Parallel Agent Reinforcement Learning, coordinates up to 100 sub agents and about 1,500 tool calls per task and reports about 4.5 times faster execution on wide search workloads. Benchmarks show strong results on SWE Bench, MMMU Pro, VideoMMMU, HLE, and BrowseComp..... Full analysis: https://lnkd.in/gFDfxbMN Model weight: https://lnkd.in/gEARyE5D? Technical details: https://lnkd.in/gEARyE5D? Try it here: https://www.kimi.com/agent Moonshot AI

Apple Researchers Release CLaRa: A Continuous Latent Reasoning Framework for Compression‑Native RAG with 16x–128x Semantic Document Compression Apple Researchers Release CLaRa-7B, a continuous latent reasoning framework that replaces raw documents with learned memory tokens and unifies retrieval and generation in a shared embedding space. A Mistral-7B backbone with LoRA adapters and SCP pretraining on ≈2M Wikipedia passages delivers 4x–128x semantic compression while improving average F1 over LLMLingua-2 by up to 17.31 points in Oracle settings and even outperforming BGE + full-text RAG, reaching 96.21 Recall@5 and 75 F1 on Natural Questions and HotpotQA at 4x compression..... Full analysis: https://lnkd.in/gkBDpmcH Paper: https://lnkd.in/gQDxgRQd Model weights on HF: https://lnkd.in/gDPRhQNb Repo: https://lnkd.in/gC3eT8ch Apple Jie He He Bai Sinead Williamson Navdeep Jaitly Yizhe Zhang

NVIDIA and Mistral AI Bring 10x Faster Inference for the Mistral 3 Family on GB200 NVL72 GPU Systems NVIDIA announced today a significant expansion of its strategic collaboration with Mistral AI. This partnership coincides with the release of the new Mistral 3 frontier open model family, marking a pivotal moment where hardware acceleration and open-source model architecture have converged to redefine performance benchmarks. This collaboration is a massive leap in inference speed: the new models now run up to 10x faster on NVIDIA GB200 NVL72 systems compared to the previous generation H200 systems. This breakthrough unlocks unprecedented efficiency for enterprise-grade AI, promising to solve the latency and cost bottlenecks that have historically plagued the large-scale deployment of reasoning models.... Full analysis: https://lnkd.in/gs7YdT6V Models on HF: https://lnkd.in/gDjisYpY Corporate Blog: https://pxllnk.co/6tyde68 Dev Blog: https://pxllnk.co/xvq4zfm NVIDIA NVIDIA AI NVIDIA Robotics NVIDIA for Startups

OceanBase open-sources seekdb: An Open Source AI Native Hybrid Search Database for Multi-model RAG and AI Agents seekdb is an AI native search database that unifies relational data, vector search, full text search, JSON and GIS in one MySQL compatible engine. It provides hybrid search through DBMS_HYBRID_SEARCH and in database AI functions such as AI_EMBED, AI_COMPLETE and AI_RERANK, so RAG and agentic applications can run retrieval and orchestration inside a single system...... Full analysis: https://lnkd.in/gFGVTrJ9 Repo: https://pxllnk.co/4orh5qr Project: https://www.oceanbase.ai/ OceanBase

Perplexity AI Releases TransferEngine and pplx garden to Run Trillion Parameter LLMs on Existing GPU Clusters How can teams run trillion parameter language models on existing mixed GPU clusters without costly new hardware or deep vendor lock in? Perplexity’s research team has released TransferEngine and the surrounding pplx garden toolkit as open source infrastructure for large language model systems. This provides a way to run models with up to 1 trillion parameters across mixed GPU clusters, without locking into a single cloud provider or buying new GB200 class hardware..... Full analysis: https://lnkd.in/gJkc6t3r Paper: https://lnkd.in/gB9-W8DF Repo: https://lnkd.in/g6-_dfsd Perplexity Nandor Licker Kevin Hu Vladimir Zaytsev Lequn Chen Aravind Srinivas

Meta AI Releases Segment Anything Model 3 (SAM 3) for Promptable Concept Segmentation in Images and Videos Meta’s Segment Anything Model 3 (SAM 3) is a 848M parameter vision foundation model that upgrades Segment Anything from promptable visual segmentation to Promptable Concept Segmentation, unifying image and video detection, segmentation and tracking from text prompts, exemplars, points and boxes. Trained and evaluated on the new SA Co stack with about 270K evaluated concepts and over 4M automatically annotated concepts, SAM 3 approaches 75–80 percent of human cgF1 and sets a new reference baseline for open vocabulary image and video segmentation.... Full analysis: https://lnkd.in/gX_yBRrw Paper: https://lnkd.in/geZsevy5 Model weights: https://lnkd.in/gxqymx89 Repo: https://lnkd.in/gV7K2iHt Meta AI at Meta Encord

Cerebras Releases MiniMax-M2-REAP-162B-A10B: A Memory Efficient Version of MiniMax-M2 for Long Context Coding Agents MiniMax-M2-REAP-162B-A10B is a Sparse Mixture-of-Experts Causal Language Model created by applying Router weighted Expert Activation Pruning, REAP, to the 230B MiniMax-M2 at a 30% expert pruning rate, resulting in 162B total parameters with 10B active per token, 62 layers, 48 heads, 180 experts and a 196,608 token context window, while maintaining near identical accuracy to MiniMax-M2 on HumanEval 93.3, MBPP 86.5, AIME25 73.3, MATH-500 89.4 and τ² bench Telecom 59.1, making it a memory efficient long context coding and tool calling model for vLLM deployments..... Full analysis: https://lnkd.in/g3wftvWQ Model weights: https://lnkd.in/gcZKEtFE Related paper: https://lnkd.in/gT6DuwR9 Cerebras

Gelato-30B-A3B: A State-of-the-Art Grounding Model for GUI Computer-Use Tasks, Surpassing Computer Grounding Models like GTA1-32B How do we teach AI agents to reliably find and click the exact on screen element we mean when we give them a simple instruction? A team of researchers from ML Foundations has introduced Gelato-30B-A3B, a state of the art grounding model for graphical user interfaces that is designed to plug into computer use agents and convert natural language instructions into reliable click locations. The model is trained on the Click 100k dataset and reaches 63.88% accuracy on ScreenSpot Pro and 69.15% on OS-World-G, with 74.65% on OS-World-G Refined. It surpasses GTA1-32B and larger vision language models such as Qwen3-VL-235B-A22B-Instruct..... Full analysis: https://lnkd.in/gKrkuqfC Model weights: https://lnkd.in/gvQyaCVh Repo: https://lnkd.in/gGCUeBfJ Dhruba Ghosh Aylin Akkus Yuhui Zhang Marianna N. Jenia Jitsev Yejin Choi Ludwig Schmidt

Memori: An open-source SQL-Native memory engine for AI Memori enables any LLM to remember conversations, learn from interactions, and maintain context across sessions with a single line: memori.enable(). Memory is stored in standard SQL databases (SQLite, PostgreSQL, MySQL) that you fully own and control. Why Memori? ✅ One-line integration - Works with OpenAI, Anthropic, LiteLLM, LangChain, and any LLM framework ✅ SQL-native storage - Portable, queryable, and auditable memory in databases you control ✅ 80-90% cost savings - No expensive vector databases required ✅ Zero vendor lock-in - Export your memory as SQLite and move anywhere ✅ Intelligent memory - Automatic entity extraction, relationship mapping, and context prioritization Repo: https://pxllnk.co/lzpp4xm #opensource #ai #artificialintelligence #aiagent #aiagents

StepFun AI Releases Step-Audio-EditX: A New Open-Source 3B LLM-Grade Audio Editing Model Excelling at Expressive and Iterative Audio Editing How can speech editing become as direct and controllable as simply rewriting a line of text? StepFun AI has open sourced Step-Audio-EditX, a 3B parameter LLM based audio model that turns expressive speech editing into a token level text like operation, instead of a waveform level signal processing task. Step-Audio-EditX reuses the Step-Audio dual codebook tokenizer. Speech is mapped into two token streams, a linguistic stream at 16.7 Hz with a 1024 entry codebook, and a semantic stream at 25 Hz with a 4096 entry codebook. Tokens are interleaved with a 2 to 3 ratio. The tokenizer keeps prosody and emotion information, so it is not fully disentangled. On top of this tokenizer, the StepFun research team builds a 3B parameter audio LLM. The model is initialized from a text LLM, then trained on a blended corpus with a 1 to 1 ratio of pure text and dual codebook audio tokens in chat style prompts. The audio LLM reads text tokens, audio tokens, or both, and always generates dual codebook audio tokens as output...... Full analysis: https://lnkd.in/g9CujMWF Paper: https://lnkd.in/gcUd6Ddz Repo: https://lnkd.in/gpCXKf5m Model weights: https://lnkd.in/gkqUc83S StepFun