Skip to main content
NewHolo4 is here. One model for screens, code, and tools.Read the announcement → Holo is a family of agent models. Hand Holo4 a toolbox and it works through whatever the task needs: it clicks and types on a screen, writes and runs code, or calls your APIs and MCP servers, often in the same run. The API is OpenAI-compatible, and every interface uses the same call: tools in, tool_calls out.

​
Models

Start with holo4-35b-a3b. It has a 262,144-token context, and its latency suits interactive loops. On the free tier, use holo3-1-35b-a3b. Move to holo4-27b for complex multi-step tasks and novel environments. The Holo3 generation stays available. Switching is a model ID change. IDs are stable: before a model is removed, its deprecation_date is set in GET /v1/models and a notice appears on this page, so pin an ID in production and check deprecation_date when you upgrade.

​
What Holo does

Prefer to try Holo without writing code? HoloTab runs it in your browser with no setup.

​
Benchmarks

Scores, and how each is measured, are in the release posts: Holo4 and Holo3.1.

​
Smaller sizes for your own hardware

Both Holo4 models are open weights, in BF16, FP8, NVFP4, and Q4 GGUF; each model card on Hugging Face states its license. Three smaller Holo3.1 checkpoints are not on the API, but they are open weights under Apache 2.0: download them and run them yourself. Holo3.1 35B is open too, with FP8, NVFP4, and GGUF builds for a Mac (M3 or newer, 36 GB) or a DGX Spark. Local inference covers memory, measured speed, and server setup with llama.cpp or vLLM.

​
Data retention

The API logs request time, model, and token counts. Prompts and responses are not stored: zero data retention by default. H runs its own models, so inputs and outputs are never shared with third parties. See the privacy policy.