447 results. Find your next favorite, with a little help from the community.
Open-source 6-billion parameter autoregressive language model by EleutherAI, trained on The Pile; released with weights as an alternative to GPT-3.
1T/32B-active MoE with 256K context, native INT4 QAT, Modified MIT license; deep-thinking variant with 200-300 sequential tool calls for agentic tasks.
48B/3B-active hybrid linear-attention MoE (Kimi Delta Attention + MLA at 3:1), 1M context, MIT license; 6x faster decode via ~75% KV cache reduction.
Meta's open-weight safety classifier for filtering harmful LLM inputs/outputs; initial release was a Llama 2 7B instruction-tuned model, followed by v2, v3, and v4.
Open-source TypeScript framework by Vinkius Labs for building secure MCP servers with a presenter layer that normalizes tool outputs for agents.
12B dense LLM co-developed with NVIDIA, 128k context, Apache 2.0; multilingual and quantization-aware, positioned as a drop-in replacement for Mistral 7B.
Notion's Model Context Protocol server exposing workspace read/write tools to LLM agents so they can query and edit pages, databases, and comments.
Microsoft's 7B dense Phi-3 SLM with 8K and 128K context variants (MIT license), benchmarked as a mid-tier textbook-trained model between Phi-3 mini and medium.
Microsoft's MCP server for browser automation via Playwright, giving AI agents deterministic control over pages, forms, and network for testing and scraping.
MCPOpen-source MCP server that audits agent decisions via signal fidelity, pattern classification, reliability scoring, and authority gates before commit.
MCPOpen-source MCP server enabling AI assistants to read tmux pane contents, run commands, and manage terminal multiplexer sessions.
Mistral's first open audio-input model family (Mini 3B and Small 24B), Apache 2.0; multilingual speech understanding, transcription, summarization and Q&A over audio.
Fast, cost-efficient Gemini 2.5 model with hybrid reasoning that lets developers toggle thinking on or off and set thinking budgets per request.
Microsoft's 4.2B multimodal variant of Phi-3 that accepts text + image input, small enough to run on-device for OCR, chart understanding, and vision-augmented chat.
Fast, low-cost Gemini 2.0 model with native multimodal output (images plus steerable multilingual text-to-speech) and built-in tool use including Search.
357B MoE with 200K context (up from 128K in 4.5), MIT license; improved coding, tool-use during inference, and agentic benchmark performance.
Smaller cost-optimized variant of GPT-5.4, available on the free tier via the Thinking feature and as an API model; ~4x pricier than GPT-5 equivalents in the OpenAI API.
1T/32B-active MoE with 256K context and MLA attention, Modified MIT license; 384-expert (8/token) agentic-focused foundation model from Moonshot.
Higher-compute variant of OpenAI's o3 reasoning model, available to ChatGPT Pro users and via API with tool access (web, files, Python) for high-reliability tasks.
Incremental Gemini 3 Pro upgrade with stronger core reasoning, verified 77.1% on ARC-AGI-2, shipped in preview across AI Studio, Antigravity, and Vertex AI.
A hosted MCP server for reading advertising performance and proposing approval-gated changes across Google, Meta, ChatGPT, TikTok, and LinkedIn Ads.
A ChatGPT Work agent that investigates connected company data, explains changes, builds interactive dashboards, and carries out user-approved follow-up actions.
A local MCP server and companion tools that measure how much coding-agent token spend becomes output instead of repeated context reads.
Meta's personal AI agent that can plan and carry out everyday tasks in a dedicated secure virtual machine while keeping sensitive actions under user approval.