---
license: other
library_name: transformers
base_model: openai/gpt-oss-120b
pipeline_tag: text-generation
tags:
- gpt-oss
- code
- agentic
- swe
- tool-use
- research-preview
- custom-code
- vllm
- mxfp4
---

# gpt-oss-160b-kiwi

`gpt-oss-160b-kiwi` is an agentic coder version of GPT-OSS 120B.

After a bunch of iterations in a recursive coding-agent harness, this is the end result of the current 160B branch. It expands the 120B base with `48` added specialist experts and is intended to run at `12` active experts per token.

This model was trained on a 2-GPU setup.

This is by far one of the best checkpoints from this project so far. The next `180B` line should be better, but Kiwi is the current strong agentic-coder release.

## Overview

- Base model: `openai/gpt-oss-120b`
- Total expert rows: `176`
- Added specialist experts: `48`
- Format: `MXFP4`
- Recommended active experts: `top-k=12`
- Intended use: coding, agentic coding, SWE-style workflows, tool-using automation
- Status: research preview

## Recommended vLLM

This model was tested with vLLM using the GPT-OSS reasoning and OpenAI tool-call parsers.

```bash
vllm serve /path/to/model \
  --served-model-name vllm/doobee \
  --tensor-parallel-size 2 \
  --max-model-len 60000 \
  --gpu-memory-utilization 0.88 \
  --enforce-eager \
  --trust-remote-code \
  --reasoning-parser openai_gptoss \
  --tool-call-parser openai \
  --enable-auto-tool-choice \
  --hf-overrides '{"num_experts_per_tok": 12}'
```

Recommended parameters:

- `num_experts_per_tok=12`
- `tensor-parallel-size=2`
- `max-model-len=60000`
- `gpu-memory-utilization=0.88`
- `reasoning-parser=openai_gptoss`
- `tool-call-parser=openai`
- `enable-auto-tool-choice`

The staged config is already set to `num_experts_per_tok=12` and `experts_per_token=12`. If your runtime ignores those fields, pass the `--hf-overrides` value explicitly.

## Tool Calling

Kiwi was primarily built and tested as an agentic coding model.

Recommended temperatures:

- `0.0` for deterministic tool use
- `0.3` for steady coding-agent work
- `0.6` for more flexible agentic exploration

The recommended serving path is OpenAI-compatible Chat Completions with vLLM's GPT-OSS reasoning parser and OpenAI tool-call parser enabled.

## Next

Kiwi is the current strong 160B agentic-coder checkpoint. The upcoming 180B line is expected to push this further.

## License

Replace the placeholder `license: other` metadata with the actual license you want to publish under after confirming compatibility with the base model and your added weights.