Models
Updated Oct 2026
A curated selection of open models — every one runs in llama.app.
Good places to start
LFM 2.5 24B-A2B 24 GB+ memory
Liquid AI's fast local model for reasoning and tool use.
Qwen 3.8 27B 32 GB+ memory
Alibaba's multimodal mid-sized models for coding and research, long-horizon agentic work with 256K native context.
Qwen 3.8 Flash Next 125B-A6B 128 GB+ memory
An experimental preview of the Qwen 4 architecture for long-context agents on large setups, with vision and 256K context.
Muse Glimmer 30B 32 GB+ memory
Meta's local model to run on your laptop powering always-on agents, with coding, tool use, and image understanding.
Gemma 4 E2BE4B12B26B-A4B31B 16 GB+ memory
Google's most capable open agentic models for your laptops, with vision, audio, and support for 140+ languages.
GPT-OSS 20B120B 24 GB+ memory
OpenAI's open-weight family for local reasoning and agents, with tool use, function calling, and adjustable reasoning effort.
Gemma 3 270M1B4B12B27B 8 GB+ memory
Google's efficient multimodal models sized from phones to workstations, with vision, 140+ languages, and up to 128K context.
More models
Nemotron 3.5 Super VL 120B-A12B 128 GB+ memory
NVIDIA's large hybrid Mamba-Transformer mixture-of-experts model with vision. Strong at reasoning, coding, and documents, with up to 1M context.
Muse Spark Open 816B-A42B 512 GB+ memory
Meta's large and capable agentic model, takes image and audio input, with adjustable reasoning, and 1M context.
Nemotron 3.5 Lightning 30B-A3B 32 GB+ memory
NVIDIA's efficient long-context model for local agents with switchable reasoning mode and up to 1M context.
Laguna XS 2.1 33B-A3B 32 GB+ memory
Poolside's compact coding agent for long-horizon work, with interleaved reasoning and 256K context.
Laguna S 2.1 118B-A8B 96 GB+ memory
Poolside's larger coding model for large software projects with 1M context.
DeepSeek V4 304B 192 GB+ memory
DeepSeek's reasoning and coding model for large setups, combining sparse MoE efficiency with 1M context.
Qwen 3.6 27B35B-A3B 32 GB+ memory
Alibaba's efficient multimodal models that punch above their size for coding, reasoning, and vision tasks.
GLM 4.7 Flash 32 GB+ memory
Z.ai's fast local agentic coding model, built for software engineering, browser automation, and multi-turn tool use.
Qwen 3.5 Small 0.8B 8 GB+ memory
Alibaba's compact multimodal reasoning models for phones and laptops, with vision, switchable thinking, and support for 201 languages.
Devstral 2 24B123B 24 GB+ memory
Mistral AI's mid-sized models for autonomous software agents, built to explore repositories, edit across files, and solve engineering tasks.
Ministral 3 3B3B Reasoning8B8B Reasoning14B14B Reasoning 8 GB+ memory
Mistral AI's compact multimodal models for on-device apps, offering strong performance across 3B, 8B, and 14B sizes.
Decision models
Models that take a document and typed questions, and return a probability for each option -- route a ticket, moderate content, pick an agent’s next step. They don’t chat; apps call them through the API. How to use ↗
Kev 0.8B4B9B 8 GB+ memory
Decision models that take context and typed questions as input, and output a calibrated probability for each option.
Clef Flash 9B 16 GB+ memory
Cloudflare's decision model. Reads a text, JSON, or image state and returns a probability for every option of every typed question in a single pass.
Laya 421M 8 GB+ memory
A multilingual decision model for documents and structured data, returning calibrated probabilities across 100+ languages.
Bespoke Nimble 9B 16 GB+ memory Non-commercial
Bespoke Labs' mid-sized decision model. Given a context and typed questions, it returns a probability for each option.
Embedding models
Models that turn text — and, for some, images and audio — into vectors, for search, retrieval, and clustering. They don’t chat; apps call them through the API. How to use ↗
Find thousands more GGUF models on Hugging Face ↗, then choose Llama under “Use this model” on a model’s page.