🦙

Open Source LLM Selector for Agents

Pick the best open-weight model for your AI agent use case. Compare Llama, Qwen, Mistral, DeepSeek, Gemma, Phi, and more by fit, VRAM, context, and speed.

Agent requirements

Primary use case Max available VRAM (GB) Minimum context (k tokens) Deployment preference

Only show models with native vision support

Estimates are directional. Last updated: 2026-07-07. See notes.

Ranked recommendations

—

Select a use case to see ranked models.

Model comparison

Model	Family	Params	Context	Q4 VRAM	Score	Best fit
Select a use case to populate the table.

Frequently asked questions

Which open-source model is best for coding agents?

DeepSeek Coder V2 and DeepSeek V3 lead open-weight coding benchmarks. Llama 3.1 405B and Qwen 2.5 72B are strong generalist alternatives with good code performance. For local hardware, Llama 3.1 70B and Qwen 2.5 32B offer the best quality/speed trade-offs.

What is the best open model for tool-using agents?

Look for models with strong function-calling scores and long context. Mistral Large 2, Llama 3.1 70B/405B, Qwen 2.5 72B, and Llama 3.1 8B (for lightweight tasks) all have robust tool support in modern inference engines.

Can I run these models locally?

Many fit consumer hardware at 4-bit quantization: Llama 3.1 8B, Llama 3.2 3B, Gemma 2 9B, Qwen 2.5 14B, Mistral NeMo 12B, and Phi 4 all run on 8-24 GB GPUs. Larger models like Llama 3.1 70B, Qwen 2.5 72B, or DeepSeek V3 need multi-GPU or cloud API hosting.

How accurate are the use-case scores?

Scores are relative (0-10) directional ratings based on public benchmarks and community reports as of mid-2026. They help narrow the field; always benchmark the final candidates on your own prompts and agent traces.

Which model is best for non-English agents?

Qwen 2.5 models excel at multilingual coverage, followed by Mistral Large 2 and Llama 3.1. For primarily English workloads, Gemma 2 and Llama 3.1 are excellent.

What about vision agents?

Qwen2-VL 72B is the strongest open vision-language model for document and image understanding. Llama 3.2 Vision 11B is a practical, consumer-GPU-friendly alternative.

VRAM Calculator → Inference Engine Selector → API Pricing Comparator → Agent Benchmark Selector →

Scores are relative (0-10) and based on public benchmarks and community reports as of mid-2026. VRAM estimates assume KV cache plus a modest batch/context reserve; real usage varies by backend and quantization. Always benchmark on your own prompts.