Model Library

Ideal for autocomplete, classification, and low-latency chat. Runs in-browser with WebGPU.

Versatile 3B model for on-device chat and reasoning. Fits comfortably in mid-range mobile VRAM.

Balanced 8B with a 128k context window. Great for document summarisation and tool-use.

Frontier-class open model. Requires multi-GPU desktop or mesh dispatch. Not browser-eligible.

Outstanding reasoning at 3.8B. 128k context window. Runs on mid-range consumer GPUs.

Microsoft's latest Phi-4 excels at STEM, coding and multi-step reasoning. Strong math performance.

Ultra-lightweight multilingual model for edge deployment. Fits on any device with a WebGPU runtime.

Strong multilingual and coding capabilities. Best run on desktop or mesh nodes.

Specialised coding variant of Qwen 2.5. Supports fill-in-middle (FIM) for IDE-style completions.

Top-tier multilingual model with 128k context. Rivals GPT-4o on many benchmarks. Mesh-dispatched.

Reliable general-purpose model with strong instruction following. Wide hardware support.

Mistral × NVIDIA 12B with 128k context. Excellent multilingual performance and function-calling.

Mixture-of-Experts with 8×7B experts, activating 13B per token. High quality at competitive speed.

Google's compact 2B model trained with knowledge distillation. Punches above its weight.

Highly capable 9B from Google DeepMind. Competitive with models twice its size on many benchmarks.

Google's most capable open Gemma model. Excellent reasoning and instruction following.

Chain-of-thought reasoning specialist distilled from DeepSeek-R1. Excels at math and coding.

Larger R1 distillation with stronger multi-step reasoning. Best for complex analysis tasks.

State-of-the-art coding MoE model. Supports 338 programming languages and 160k context.

Optimised for retrieval-augmented generation (RAG). Grounding, citations, and tool-use built in.

Lightweight code generation with fill-in-middle. Trained on 600+ languages. Browser-eligible.

High-quality 15B code model with FIM. Best performance on competitive programming tasks.

High-quality text embeddings at near-zero cost. Runs in any browser, sub-5ms per batch.

State-of-the-art multilingual embedding model (100+ languages). Dense + sparse + ColBERT in one.

Instruction-tuned LLM repurposed for high-quality long-context embeddings. Best for RAG pipelines.

Visual instruction-tuned model. Accepts images + text. Suitable for document and chart analysis.

Tiny vision-language model designed for edge deployment. Runs in-browser for image Q&A.