Models that run on your device

Meta's new agent-first model - built for reliable tool use, working in project folders, and recovering from its own mistakes. Sees images too. Brand-new: early support, expect rough edges.

Alibaba's newest. Frontier coding, math, and reasoning with built-in thinking; a hybrid attention design keeps long documents fast. The strongest model here for 24GB-class graphics cards.

Top-tier coding, math, and reasoning. The 35B MoE runs fast for its size — only 3B parameters active per token.

Google's latest open model. Strong writing, analysis, and instruction-following. E2B/E4B run on modest machines; the 26B is a fast MoE.

Mistral's latest small model family. Fast inference, great for on-device use.

Microsoft's compact model. Best option for machines with limited RAM.

OpenAI's open-weight model. Strong reasoning, agentic tasks, and function calling.

Liquid AI's small mixture-of-experts - 8B total, 1.5B active per token, so it runs fast on ordinary machines. Strong instruction following and tool use for its size; 10 languages.

IBM's small hybrid mixture-of-experts - 7B total, about 1B active per token, with a 1M-token context. Apache-2.0. A fast, permissive everyday model for long documents.

NVIDIA's newest open reasoning model - a 30B mixture-of-experts with 3B active per token and a 256K context. Strong reasoning, coding, and tool use; runs on a 32 GB machine with the experts in main memory.

Ant Group's mid-size mixture-of-experts - 16B total, 1.4B active per token. MIT-licensed, 128K context; a quick all-rounder for 24 GB machines.

DeepSeek's V4 Flash - a 284B mixture-of-experts with 13B active per token and a 1M-token context. Frontier-class reasoning and agentic work on a workstation with a big graphics card and 128 GB or more of main memory.

Zhipu's GLM-5.2 - a 753B mixture-of-experts with a 1M-token context. For very large workstations: a big graphics card and 384 GB or more of main memory, experts in main memory.

Powerful chain-of-thought reasoning. Excels at complex problem solving, math, and logic.

Mistral's agentic coding model. Tops open models on SWE-bench at its size — purpose-built for software engineering and code agents.

DeepReinforce's agentic coding model, self-improved with RL - 1.5 extends the loop to generating its own training tasks. State-of-the-art among open coders at its size, purpose-built for terminal coding agents and tool use. The 9B runs on modest machines; the 35B is a mixture-of-experts (3B active per token) that runs fast for its size and adds vision.

Zhipu's fast GLM model — a strong all-rounder tuned for agentic coding, reasoning, and tool use. The lighter "Flash" tier of the GLM family; the 30B MoE runs only 3B parameters per token.

Alibaba's agentic coding model — a 30B MoE with only 3B active per token. Tuned for repository-scale work and tool use (Qwen Code, Cline).

Qwen 3.5 distilled from Claude Opus reasoning traces. Enhanced chain-of-thought capabilities.

Qwen 3.5 with safety guardrails removed. No content filtering or refusals.

Google's open medical models - discuss your own health records, lab results, and medical images (X-rays, skin photos, scans) privately on your device. For understanding and preparing questions, not diagnosis.

Empero's full-parameter distillations of the Qwen 3.8 flagship into small, fast sizes - a big knowledge jump over same-size models (the 9B scores near the giants on broad knowledge), with step-by-step reasoning. Text only.

The Qwen 3.8 flagship with refusal behavior substantially reduced (openly documented - reduced, not eliminated). Same frontier coding, math, and reasoning, and it can see images with its vision add-on. For 24GB-class graphics cards.

A community Qwen 3.5-based merge — uncensored and multimodal (it can see images you attach), with step-by-step reasoning and tool use. A creative, unfiltered generalist. v2 trains out the repetition loops of the original.