Nous-Preview API & Inference Endpoint

Design Goals

Long-horizon reasoning
Reliable enterprise tool use
Multi-domain orchestration
Failure recovery under real-world noise
Memory-efficient multi-turn planning
Verifiable execution

Results — AutomationBench (60-task subset, self-hosted)

Table with columns: Model, Size, Partial credit, Strict pass
Model	Size	Partial credit	Strict pass
Ornith-1.0-35B (base)	35B-A3B	28.9%	6.7%
seul-preview	9B	28%	7%
Qwen3.6-27B	27B	34%	10%
nous-preview	35B-A3B	37.1%	8.3%

RLVR training raises partial credit +8.2 points over the base and surpasses a dense 27B model on the same tasks while activating ~9× fewer parameters per token. Held-out environment validation improved +28% relative over the base.

Serving

bash
python -m sglang.launch_server \
  --model-path beyoru/Nous-Preview \
  --tp 4 --tool-call-parser qwen3_coder --trust-remote-code

Limitations

Trained for roughly half of the planned schedule (compute constraints); the hardest curriculum stage is still pending.
Benchmark numbers are on a 60-task public subset — strict-pass differences of 1–2 tasks are within noise.

Citation

bibtex
@misc{nous2026,
  title        = {nous-preview},
  author       = {beyoru},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/beyoru/Nous-Preview}}
}

Design Goals

Long-horizon reasoning
Reliable enterprise tool use
Multi-domain orchestration
Failure recovery under real-world noise
Memory-efficient multi-turn planning
Verifiable execution

Results — AutomationBench (60-task subset, self-hosted)

Table with columns: Model, Size, Partial credit, Strict pass
Model	Size	Partial credit	Strict pass
Ornith-1.0-35B (base)	35B-A3B	28.9%	6.7%
seul-preview	9B	28%	7%
Qwen3.6-27B	27B	34%	10%
nous-preview	35B-A3B	37.1%	8.3%

Serving

bash
python -m sglang.launch_server \
  --model-path beyoru/Nous-Preview \
  --tp 4 --tool-call-parser qwen3_coder --trust-remote-code

Limitations

Trained for roughly half of the planned schedule (compute constraints); the hardest curriculum stage is still pending.
Benchmark numbers are on a 60-task public subset — strict-pass differences of 1–2 tasks are within noise.

Citation

bibtex
@misc{nous2026,
  title        = {nous-preview},
  author       = {beyoru},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/beyoru/Nous-Preview}}
}

Nous-Preview

README

Design Goals

Results — AutomationBench (60-task subset, self-hosted)

Serving

Limitations

Citation

Explore FriendliAI today

README

Design Goals

Results — AutomationBench (60-task subset, self-hosted)

Serving

Limitations

Citation