Ornith 1.5 brings 9B reasoning to NVIDIA’s NVFP4

The Ornith model card has crossed 850,000 downloads on Hugging Face. The checkpoint combines a 9-billion-parameter dense reasoning model with NVIDIA’s NVFP4 format, targeting coding and agent workloads on a single Blackwell GPU. Hugging Face’s counter tracks file requests, so one user or automated job can generate multiple downloads.

A small model with an unusual training loop

According to the authors, Ornith 1.5 extends the training process used for Ornith 1.0, whose development drew on Qwen3.5 and Gemma 4 alongside continued pretraining, mid-training, and post-training. The new pipeline generates and selects tasks, adjusts the agent scaffolds that connect the model to tools, and improves the resulting solution rollouts through reinforcement learning.

• Architecture: 9B dense model
• Primary workloads: reasoning, coding, web research, and tool use
• Native context: 262,144 tokens
• Quantization: NVIDIA NVFP4
• License: MIT
• Other deployment target: a separate Ornith-1.5-9B-Mobile variant for mobile hardware

At BF16 precision, 9 billion parameters require about 18 GB for weights alone, with the release reporting a footprint near 19 GB. Runtime memory rises with the KV cache, batching, and framework overhead, especially near the advertised context limit. An 80 GB accelerator leaves substantial operating room, while smaller GPUs require tighter limits on context length and concurrency.

NVFP4 changes the memory math

NVFP4 stores weights as 4-bit floating-point values. Groups of 16 values share an FP8 scale, while an additional FP32 scale captures the tensor’s overall magnitude. This hierarchy preserves more range than a simple 4-bit integer representation while keeping the weight files compact.