---

base_model: open-thoughts/OpenThinkerAgent-32B
model_name: eewer/OpenThinkerAgent-32B-AWQ-Int4-Terminus2
library_name: transformers
pipeline_tag: text-generation
tags:
- qwen3
- openthinker-agent
- awq
- 4-bit
- areal-teacher
---

# OpenThinkerAgent-32B AWQ Int4

This repository contains an AWQ quantized checkpoint of
`open-thoughts/OpenThinkerAgent-32B` prepared for OPD/KDRL teacher-logprob use in AReaL.

## Quantization

- Source model: `open-thoughts/OpenThinkerAgent-32B`
- Source revision: `65d8a62b87c8d3d34bc45108a7ad87635318db9f`
- Dataset: `/wbl-fast/usrs/ee/clean-20260619/terminal-agent-rl/areal_runs/terminal-agent-demo/data/skill_based_medium.even.terminus2.slime_messages.jsonl`
- Calibration samples: 128
- Calibration seed: 7
- Max calibration token window per sample: 2048
- AutoAWQ `max_calib_seq_len`: 2048
- AutoAWQ `n_parallel_calib_samples`: 1
- AutoAWQ `max_chunk_memory`: 256 MiB
- Quantization: W4A16, group size 128, zero point True, version `GEMM`
- Modules left unquantized: `lm_head`
- `duo_scaling`: True
- `apply_clip`: True
- Torch: `2.11.0+cu128`
- CUDA: `12.8`
- AutoAWQ: `0.2.9`
- Started: `2026-06-24T22:09:24.999975+00:00`
- Finished: `2026-06-25T01:31:16.596143+00:00`
- Elapsed seconds: `12111.596`

Calibration text is rendered from the Terminus-2 medium SFT messages with the
Qwen3 chat template and `enable_thinking=True`, so thinking spans are preserved.
The quantization keeps `lm_head` unquantized for quality.

## AReaL Teacher Config

```yaml
teacher:
  path: <this checkpoint>
  quantization_config:
    method: awq
    bits: 4
    group_size: 128
    zero_point: true
    version: gemm
```

The AReaL worker environment needs `autoawq` importable only when this quantized
teacher path is used.