David Hendrickson

OpenViking has hit GitHub Trending 🏆 10k+ ⭐ in just 1.5 months since open-sourcing! Huge thanks to all contributors, users, and supporters. We’re building solid infra for the Context/Memory layer in the AI era. OpenViking will keep powering @OpenClaw and more Agent projects🚢🦞

Mistral has released Mistral Small 4, an open weights model with hybrid reasoning and image input, scoring 27 on the Artificial Analysis Intelligence Index @MistralAI's Small 4 is a 119B mixture-of-experts model with 6.5B active parameters per token, supporting both reasoning and non-reasoning modes. In reasoning mode, Mistral Small 4 scores 27 on the Artificial Analysis Intelligence Index, a 12-point improvement from Small 3.2 (15) and now among the most intelligent models Mistral has released, surpassing Mistral Large 3 (23) and matching the proprietary Magistral Medium 1.2 (27). However, it lags open weights peers with similar total parameter counts such as gpt-oss-120B (high, 33), NVIDIA Nemotron 3 Super 120B A12B (Reasoning, 36), and Qwen3.5 122B A10B (Reasoning, 42). Key takeaways: ➤ Reasoning and non-reasoning modes in a single model: Mistral Small 4 supports configurable hybrid reasoning with reasoning and non-reasoning modes, rather than the separate reasoning variants Mistral has released previously with their Magistral models. In reasoning mode, the model scores 27 on the Artificial Analysis Intelligence Index. In non-reasoning mode, the model scores 19, a 4-point improvement from its predecessor Mistral Small 3.2 (15) ➤ More token efficient than peers of similar size: At ~52M output tokens, Mistral Small 4 (Reasoning) uses fewer tokens to run the Artificial Analysis Intelligence Index compared to reasoning models such as gpt-oss-120B (high, ~78M), NVIDIA Nemotron 3 Super 120B A12B (Reasoning, ~110M), and Qwen3.5 122B A10B (Reasoning, ~91M). In non-reasoning mode, the model uses ~4M output tokens ➤ Native support for image input: Mistral Small 4 is a multimodal model, accepting image input as well as text. On our multimodal evaluation, MMMU-Pro, Mistral Small 4 (Reasoning) scores 57%, ahead of Mistral Large 3 (56%) but behind Qwen3.5 122B A10B (Reasoning, 75%). Neither gpt-oss-120B nor NVIDIA Nemotron 3 Super 120B A12B support image input. All models support text output only ➤ Improvement in real-world agentic tasks: Mistral Small 4 scores an Elo of 871 on GDPval-AA, our evaluation based on OpenAI's GDPval dataset that tests models on real-world tasks across 44 occupations and 9 major industries, with models producing deliverables such as documents, spreadsheets, and diagrams in an agentic loop. This is more than double the Elo of Small 3.2 (339) and close to Mistral Large 3 (880), but behind gpt-oss-120B (high, 962), NVIDIA Nemotron 3 Super 120B A12B (Reasoning, 1021), and Qwen3.5 122B A10B (Reasoning, 1130) ➤ Lower hallucination rate than peer models of similar size: Mistral Small 4 scores -30 on AA-Omniscience, our evaluation of knowledge reliability and hallucination, where scores range from -100 to 100 (higher is better) and a negative score indicates more incorrect than correct answers. Mistral Small 4 scores ahead of gpt-oss-120B (high, -50), Qwen3.5 122B A10B (Reasoning, -40), and NVIDIA Nemotron 3 Super 120B A12B (Reasoning, -42) Key model details: ➤ Context window: 256K tokens (up from 128K on Small 3.2) ➤ Pricing: $0.15/$0.6 per 1M input/output tokens ➤ Availability: Mistral first-party API only. At native FP8 precision, Mistral Small 4's 119B parameters require ~119GB to self-host the weights (more than the 80GB of HBM3 memory on a single NVIDIA H100) ➤ Modality: Image and text input with text output only ➤ Licensing: Apache 2.0 license

🤖 Why don't people realize you can give local AI live internet access for free? You can take an offline, quantized model (like a Qwen 9B Unsloth Q4_K_M) running locally in @lmstudio, and instantly give it the ability to: 🌐 Search the live web 🛒 Shop and compare real-time prices 📰 Retrieve breaking news The Secret: The Model Context Protocol (MCP). Just sign up for a free @brave Search account and plug their MCP server directly into LM Studio. Here is exactly how to set it up: 👀👇

Just saw this GitHub project 🛡️ OpenViking is skyrocketing 📈. This could be the best memory manager for @openclaw! 👀 ✅ OpenViking (volcengine/OpenViking) is an open-source project released by ByteDance’s cloud division, Volcengine. It's exploding in popularity and could become the standard for agentic memory. The community is already building direct plugins to integrate it with OpenClaw. Here is what I found about OpenViking as the ultimate memory manager for autonomous agents. 👇 🦞 What is OpenViking? Currently, most AI agents (like OpenClaw) use traditional RAG for memory. Traditional RAG dumps all your files, code, and memories into a massive, flat pool of vector embeddings. This is inefficient, expensive, sometimes slow, and can cause the AI to hallucinate or lose context. OpenViking replaces this. The authors call this new memory a "Context Database" that treats AI memory like a computer file system. Instead of a flat pool of data, all of an agent's memories, resources, and skills are organized into a clean, hierarchical folder structure using a custom protocol. 🚀 Why is this useful for OpenClaw? 🗂️ The Virtual File System Paradigm Instead of inefficiently searching a massive database, OpenClaw can now navigate its own memory exactly like a human navigates a Mac or PC. It can use terminal-like commands to ls (list contents), find (search), and tree (view folder structures) inside its own brain. If it needs a specific project file, it knows exactly which folder to look in (e.g., viking://resources/project-context/). 📉 Tiered Context Loading (Massive Token Savings) Stuffing massive documents into an AI's context window is expensive and slows the agent down. OpenViking solves this with an ingenious L0/L1/L2 tiered loading system: L0 (Abstract): A tiny 100-token summary of a file[5]. L1 (Overview): A 2k-token structural overview[5]. L2 (Detail): The full, massive document[5]. The agent browses the L0 and L1 summaries first. It only "downloads" the massive L2 file into its context window if it absolutely needs it, slashing token costs and API bills. 🎯 Directory Recursive Retrieval Traditional vector databases struggle with complex queries because they only search for keyphrases. OpenViking uses a hybrid approach. It first uses semantic search to find the correct folder. Once inside the folder, it drills down recursively into subdirectories to find the exact file. This drastically improves the AI's accuracy and eliminates "lost in the middle" context failures. 🧠 Self-Evolving and Persistent Memory When you close a normal AI chat, it forgets everything. OpenViking has a built-in memory self-iteration loop. At the end of every OpenClaw session, the system automatically analyzes the task results and updates the agent's persistent memory folders. It remembers your coding preferences, its past mistakes, and how to use specific tools for the next time you turn it on. 👁️ The End of the "Black Box" Developers hate traditional RAG because when the AI pulls the wrong file, it's impossible to know why. OpenViking makes the agent's memory completely observable. You can view the exact "Retrieval Trajectory" to see which folders the agent clicked on and why it made the decision it did, which I find the most useful feature. 🎯 The Bottom Line OpenViking is the missing piece of the puzzle for local autonomous AI. By giving OpenClaw a structured, file-based memory system that saves tokens and permanently learns from its mistakes, ByteDance has just given the 🦞 Clawdbots an enterprise-grade brain for free.

Holy 💩 check out the @openclaw 📊 daily PinchBench benchmark leaderboard! As more runs are amassed, averages go up or down. Qwen3.5 went up and is currently the leader 🎯. My Clawdbots are running very well on local Qwen3.5-27B unsloth Q4. 📊 Qwen 3.5 27B: 90.0% 📊 Qwen 3.5 397B-A17B: 89.1% 📊 Claude Sonnet 4.5: 88.2%