Models
The Holo family: action models trained to see interfaces and act on instructions.
Explore the Holo models available through our API and the open checkpoints released by the lab.
Holo models behind our products and API
Compare the models in our API, including Holo4 and earlier Holo releases. Each card lists the model's specifications, license, and access options.
Holo4 27B
Best accuracy on long, multi-step tasks across web, desktop, and mobile.
Model ID
holo4-27b
Architecture
Dense, 27B
Context length
262,144 tokens
Weights license
Research only (non-commercial)
Access
Paid tier, with higher rate limits
OSWorld
85.2%
Holo4 35B-A3B
Close to 27B accuracy, cheaper and faster. Open weights under Apache 2.0.
Model ID
holo4-35b-a3b
Architecture
MoE, 35B total, 3B active
Context length
262,144 tokens
Weights license
Apache 2.0 (fully open)
Access
Paid tier, with higher rate limits
OSWorld
80.8%
Holo3.1 35B-A3B
Previous-generation MoE with native function calling, on the free and paid tiers.
Model ID
holo3-1-35b-a3b
Architecture
MoE, 35B total, 3B active
Context length
65,536 tokens
Weights license
Apache 2.0 (fully open)
Access
Free tier (10 RPM) and paid tier
Holo3 122B-A10B
Previous-generation large MoE for long multi-step tasks across web, desktop, and mobile.
Model ID
holo3-122b-a10b
Architecture
MoE, 122B total, 10B active
Context length
65,536 tokens
Weights license
Research only (non-commercial)
Access
Paid tier, with higher rate limits
Open checkpoints from the lab
Documented releases on Hugging Face, with notes covering the training recipe and benchmark results where available. To run Holo4 or Holo3.1 on your own machine, see local inference in the docs.
Holo4
2026 · Apache 2.0 (35B-A3B), research only (27B)
Action models at 27B dense and 35B-A3B that click, code, and call tools across desktop, web, and mobile, in BF16, FP8, NVFP4, and GGUF.
NeoMME
2026 · Apache 2.0
Multimodal encoders and retrievers for long-context document understanding, released at 260M and 800M parameters with dense and late interaction retriever variants.
Holo3.1
2026 · Apache 2.0
Open-weight action models at 0.8B, 4B, 9B, and 35B-A3B, with quantized checkpoints for local and single-GPU deployment.
Holotron 3 Nano
2026 · NVIDIA Open Model License
A computer-use model trained on NVIDIA's Nemotron 3 Nano Omni backbone, part of H's work on efficient open backbones.
Holo3
2026 · Apache 2.0
The third generation of Holo action models, trained for grounding and multi-step control on web and desktop interfaces.
Holotron-12B
2026 · NVIDIA Open Model License
A 12B computer-use model built on NVIDIA Nemotron Nano V2 VL, released as part of H's work inside the Nemotron ecosystem.
Holo2-235B-A22B preview
2026 · CC BY-NC 4.0
A research preview of the largest Holo2 checkpoint, a mixture of experts model with 235B total and 22B active parameters.
Holo2
2025 · CC BY-NC 4.0
Holo2 at 4B, 8B, and 30B-A3B, trained for grounding and action prediction across web, desktop, and mobile screens.
Holo1.5
2025 · CC BY-NC 4.0
Open foundation models for computer-use agents at 3B, 7B, and 72B, focused on interface grounding and screen question answering.
Continue from here
Build with Holo models
Pricing, rate limits, and quickstart guides live on H Platform. Weights and quantized checkpoints live on Hugging Face.