Built for thoughtful discussion and helpful, structured reasoning. Strong multilingual generation, precise instruction-following, and clean formatting across technical and analytical tasks.

The newest Hermes generation in an efficient 70B class. Balanced speed and reasoning depth with controlled verbosity, reliable structured output, and strong code generation and debugging.

Phi-Reasoning delivers long, thorough reasoning responses for deep research and complex problem solving. Optimal for in-depth analysis, academic assistance, and technical explanations.

A 27B dense hybrid model pairing Gated Delta Networks with attention for long-context efficiency. Strong at coding, reasoning, and agent tasks across 201 languages under Apache 2.0.

A 12B collaboration between Mistral and NVIDIA, designed to be a drop-in efficient workhorse. State-of-the-art multilingual coverage, large context, and strong instruction following on a single GPU.

A 671B MoE flagship activating 37B per token. Exceptional math and reasoning with MIT-licensed weights. A community benchmark king that rivals closed frontier labs at a fraction of cost.

Poolside's leading coding agent model for intricate software development and agentic engineering workflows. Excels at repository-scale reasoning, debugging, refactoring, and autonomous development.

An advanced model combining deep engineering productivity with speed and emotional intelligence. Excels at autonomously driving end-to-end project execution, debugging, and log analysis.

A fast, next-generation model combining engineering productivity with speed. Stands out at autonomously driving end-to-end project execution and handling complex log analysis and debugging.

A cheap MoE reasoning model designed for coding and agent performance. Dynamically routes tokens through specialized experts with a hybrid attention architecture for fast inference.

The open coding workhorse that matches GPT-4o on code benchmarks while running locally. Supports 92 programming languages, fill-in-the-middle completion, and long repository context.

A MoE code specialist activating 21B of 236B parameters. State-of-the-art on HumanEval and competitive programming, with strong repo-level understanding and an MIT-licensed Lite variant for local rigs.

A unified iteration consolidating several Mistral models into one system: Magistral's reasoning, Pixtral's multimodal understanding, and Devstral's agentic coding in one workflow.

A cost-effective multimodal foundation model that efficiently processes text, images, and video, prioritizing speed and affordability with a massive 1M-token context window.

A leading open vision-language model with precise document parsing, object grounding, and long-video understanding. Excels at OCR, chart analysis, and UI agent control.

A natively multimodal 12B model with a 400M vision encoder. Handles variable image sizes and interleaved text+image at high throughput, runnable on a single 24GB consumer GPU.

Meta's flagship open vision model for high-resolution image reasoning, document QA, and visual grounding. Built on the proven Llama 3 backbone with an adapter-based image encoder.

A top open-weight vision specialist closing the gap with GPT-4o on multimodal benchmarks. Outstanding at fine-grained perception, multi-image reasoning, and scientific diagram understanding.

The de-facto open speech-to-text standard. Robust multilingual transcription and translation across 99 languages with strong noise resilience. Runs in real time via whisper.cpp on consumer hardware.

A versatile audio-language model that understands speech, ambient sound, and music. Supports voice-chat and audio-analysis modes for transcription, captioning, and spoken Q&A.

A versatile embedding model supporting dense, sparse, and multi-vector retrieval in one. Handles 100+ languages and inputs up to 8192 tokens, the go-to backbone for local RAG pipelines.

A highly-tuned 137M parameter local text embedding model. Features a massive 8192 context window and support for variable output dimensions, perfect for local RAG pipelines.

A non-autoregressive speech synthesis model utilizing flow matching. Capable of high-fidelity, zero-shot voice cloning with highly natural prosody and voice inflections.

An experimental real-time spoken audio-to-audio foundation model. Processes speech and generates natural, low-latency voice responses natively.

A compact, efficiency-first assistant model tuned for fast everyday reasoning, summarization, and chat. Designed to deliver responsive performance on modest hardware while staying lightweight.

A radically efficient 1.58-bit ternary-weight architecture. Matches full-precision quality at a fraction of the memory and energy, hinting at CPU-only LLM inference. Early research weights.

A high-efficiency State Space Model (SSM) architecture. Boasts linear time complexity and massive context capabilities, matching Transformers with significantly less overhead.

An RNN architecture designed to deliver Transformer-level quality. Featuring linear attention and low-resource multilingual execution without traditional scaling penalties.

A hybrid architecture combining linear attention State Space Models with Mixture-of-Experts (MoE) to handle large context windows with exceptional generation speeds.

A native early-stage mixed-modal architecture engineered to ingest and output both images and text sequentially, rather than using separate external projection encoders.