---
license: apache-2.0
language:
- en
base_model: Qwen/Qwen3.5-0.8B-Base
pipeline_tag: zero-shot-classification
library_name: transformers
tags:
- ollaya
- sura
- decision
- classification
- distillation
- qwen3_5

---

# Sura 0.8b

> **In Ollaya's registry as `sura:0.8b`.** Ollaya's runtime matches upstream `pplx_decider` on all 696 parity
> questions (CPU and CUDA). On Ollaya Bench 2026.10 it scores 37.2 (typed-decisions 0.560, public benchmark macro
> accuracy 0.586): level with coba:1b, below decider:0.8b and kev:0.8b. [`sura-2b`](https://huggingface.co/ollaya-dev/sura-2b)
> is the more accurate tag.

Sura 0.8b is a decision model by [Ollaya](https://ollaya.dev). The smaller sura, aimed at the best accuracy under 1B parameters. Give it a **state** (a message, a support ticket,
a JSON object, an agent's next tool call) and **questions**, and it answers every question in one forward pass with
calibrated probabilities:

- `noul`: yes or no, as the probability of yes;
- `choice`: one of up to 255 options, each described in plain words;
- `score`: a level on a scale of described levels.

## Ollaya's model families

| Family | Network | Built for |
|---|---|---|
| [coba](https://huggingface.co/ollaya-dev/coba-400m) | encoder (Ettin), 68M to 1B | speed: CPUs, and the browser at 35 to 147 MB |
| **sura** | decoder (Qwen3.5) | accuracy |

| Model | Base | Status |
|---|---|---|
| [`sura-0.8b`](https://huggingface.co/ollaya-dev/sura-0.8b) | `Qwen/Qwen3.5-0.8B-Base` | `sura:0.8b`; Ollaya Bench 37.2 |
| [`sura-2b`](https://huggingface.co/ollaya-dev/sura-2b) | `Qwen/Qwen3.5-2B-Base` | `sura:2b` and `sura:latest`; Ollaya Bench 40.9 |

## How it is built

- **Network**: the joint inline-readout protocol of Perplexity's pplx-decider (version 3). Every question of a
  request shares one sequence with the state; each question is followed by its options and a marker, and a
  255-way linear readout scores the options at that marker. The full-attention layers see the whole sequence;
  the Gated DeltaNet layers stay causal.
- **Teacher**: [`perplexity-ai/pplx-decider-flash-v1.2`](https://huggingface.co/perplexity-ai/pplx-decider-flash-v1.2)
  (Qwen3.5-4B, Apache-2.0). The student shares its tokenizer and prompt, so it learns the teacher's full
  distribution over the options, not only its answers.
- **Data**: the mixture of coba's zero-shot models, about 340,000 states with gold and teacher labels. It holds
  none of the datasets of the benchmarks Ollaya reports, so the results are zero-shot.
- **License**: Apache-2.0, like the base model and the teacher.

## Run it

With [Ollaya](https://ollaya.dev) (Linux, macOS, Windows; CPU, CUDA, Metal):

```sh
curl -fsSL https://ollaya.dev/install.sh | sh
ollaya run sura:0.8b --ask "Is the customer asking for a refund?" < ticket.txt
```

## Files

- `model.safetensors` (BF16 backbone) and `readout.safetensors` (the 255-code readout).
- The teacher's processor files (tokenizer, chat template, `joint_config.json`), so upstream's `pplx_decider` loads
  this directory as a pplx-decider checkpoint.
- `source/`: Perplexity's `pplx_decider` package from pplx-decider-flash-v1.2, unchanged (Apache-2.0, see `NOTICE`).