Small Models.
Serious Capability.

Open-weight language, code and vision-language models from 494M to 4.5B parameters — trained on signal-dense data, checked for behavior, and shipped in formats that run on the hardware you already own.

The Pluto AI campus in 3D: the lab headquarters, glowing data lines, cooling towers and compute cabinets

Our Models

Atlas-Coder-2 data towers in the Pluto AI campus: a server rack tower, a standing graphics card and a stacked-memory tower

Atlas-Coder-2-0.5B — a top-5 sub-1B code model on EvalPlus

  • 494M Parameters
  • Qwen2.5-Coder-0.5B-Instruct Base
  • EvalPlus Top 5 Sub-1B
  • 50K Execution-Verified Samples
  • QLoRA · One Kaggle Session
  • 75 Tokens Per Second · M2 Air
  • Safetensors · GGUF
  • Apache-2.0 License
Cooling towers and GPU compute cabinets in the Pluto AI campus

Apollo-VL-Edge-3B — top-3 sub-5B vision-language model for charts, documents and OCR

  • 3.0B Parameters · ~6 GB In FP16
  • ChartQA 78.6%
  • AI2D 77.98%
  • OCRBench 786
  • 161,562 Training Examples
  • GGUF · MLX 4/6/8-Bit
  • Apache-2.0 License
The Pluto research lab building with its ringed observatory dome

Athena-1.0-4B — document and chart reasoning in 4.5B parameters

  • 4.54B Parameters
  • Qwen3.5-4B Base
  • 250,000 Training Rows
  • Deduplicated & Decontaminated Data
  • One H100 · ~18 Hours
  • BF16 Safetensors · GGUF
  • Apache-2.0 License
The glowing Pluto planet mark at the end of the data line

Pluto-Genesis-0.6B — instructions, math and code under 1B parameters

  • 596M Parameters
  • Qwen3-0.6B Base
  • 80K Curated Samples
  • Instructions · Math · Code
  • QLoRA · Single T4
  • Research Paper On Zenodo
  • Safetensors · GGUF
  • Apache-2.0 License

Pull a model tonight. Run it offline tomorrow.

All models on Hugging Face