gpt-oss-140b Ren-2
gpt-oss-140b, codename Ren-2, is an agentic / SWE-oriented derivative of OpenAI GPT-OSS 120B.
This release takes the 120B base model and adds roughly 20B more parameters oriented toward agentic coding and SWE-style behavior.
Overview
- Base model:
openai/gpt-oss-120b
- Release name:
gpt-oss-140b Ren-2
- Format:
MXFP4
- Intended use: coding, agentic coding, SWE-style assistant workflows
- Status: research preview
Ren-2 is meant to feel like a more agentic version of GPT-OSS 120B rather than a generic continuation of the base checkpoint.
Training
- Built on a custom framework
- Roughly
3 hours of pre-training / post-training work for this release path
- Expanded from the 120B base with roughly
20B additional parameters
This is an iterative open release. More sizes and follow-up revisions will come later.
Inference
This model was tested with:
Recommended serving settings:
num_experts_per_tok=12--reasoning-parser openai_gptoss--tool-call-parser openai--enable-auto-tool-choice
The original base setup used 4 active experts per token. Ren-2 is intended to run at 12 active experts per token.
Rough active-equivalent compute:
- original
top-k=4: about 5.7B active-equivalent params
- Ren-2
top-k=12: about 12.9B active-equivalent params
These are approximate active-equivalent numbers, not total parameter counts.
In internal agentic-task traces at top-k=12, roughly half of active routing traffic ran through the added 20B expansion. Observed new-expert usage in those traces was about 48.6% of active expert selections and about 46.2% of routing mass. This is workload-dependent.
This release is intended to run directly from the baked model shards. No extra router merge step is required at inference time.
What It Is Good At
- coding
- agentic coding
- SWE-style assistant behavior
- practical tool-using workflows
Ren-2 is intended to be usable for production-style coding and agentic workflows, including terminal coding agents, SWE assistants, and tool-using automation setups.
Feedback
Useful feedback includes:
- coding quality
- tool use quality
- long-context behavior
- inference stability
- preferred smaller sizes / VRAM targets
If you want smaller custom models, reach out with the hardware target and the kind of feedback you can provide.
It can be a different size or architecture, as long as the feedback loop is useful.
Included Files
config.jsongeneration_config.jsontokenizer.jsontokenizer_config.jsonchat_template.jinjamodel.safetensors.index.jsonmodel-*.safetensorsREADME.md
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.
