File size: 2,344 Bytes
da9d400 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 | ---
license: other
license_name: genysis-iq-commercial
license_link: LICENSE
library_name: mlx
pipeline_tag: image-text-to-text
tags:
- anura
- genysis-iq
- agentic
- coding
- mlx
- vision
- tool-use
---
# Anura R1 Ultra
**Anura R1 Ultra** is an agentic coding model by **Genysis IQ** (version 1.0.0).
- Agentic software engineering: reads and edits code, runs commands and tests, plans multi-step work
- Web search and crawling for current information, with cited sources
- Vision: screenshots, diagrams, UI mockups, charts and photos
- Fast local inference on Apple Silicon with MLX (mixture-of-experts, ~3B active parameters per token)
- 128K-token context in the default engine configuration
## Results (measured on an Apple M5, 32 GB)
| | Anura R1 Ultra |
|---|---|
| Identity, held-out questions (Anura system prompt / served with identity preamble) | 26/26 · 26/26 |
| Identity, bare weights under a generic third-party prompt | 24/26 |
| Vision (shapes, OCR, code screenshot, chart, UI→HTML) | 5/5 |
| Agent coding tasks, pass@1, 40 verified tasks in 6 languages | 36/40 |
| Decode speed | ~48 tok/s |
| Cached follow-up prefill (14.7K-token prompt) | 0.18 s |
The coding tasks are the suite the training trajectories came from, so they show the fine-tune kept the
foundation's skills; they are not a held-out benchmark.
## Format
MLX 4-bit weights plus the Anura R1 Ultra LoRA adapter in `anura_adapter/`, applied at load time. The chat
template is Anura R1 Ultra's own (`chat_template.jinja`). About 20 GB of memory is needed (32 GB Apple
Silicon Mac recommended).
## Running it
The intended runtime is the Anura engine and CLI from Genysis IQ (`anura serve`, `anura`), which adds the
agent tools, sessions, web UI and an OpenAI-compatible API. To load the weights directly with mlx-vlm:
```python
from mlx_vlm import load, generate
model, processor = load("<this repo>", adapter_path="<this repo>/anura_adapter")
```
## License and provenance
Access to this repository is limited to licensed users of Genysis IQ. Genysis IQ built Anura R1 Ultra by
fine-tuning an open-weight foundation model released under the Apache License 2.0; those foundation weights
remain under that license, and `NOTICE` gives the attribution it requires. Genysis IQ's adapter and packaging
are provided under Genysis IQ's commercial terms.
|