DEV.co
Open-Source LLM · unsloth

gpt-oss-120b-GGUF

gpt-oss-120b is OpenAI's open-weight 120B parameter model (5.1B active via MoE) released under Apache 2.0. It is a GGUF-quantized version maintained by Unsloth, designed for reasoning-heavy tasks, agentic workflows, and general-purpose production use. The model uses native MXFP4 quantization in its MoE layer and requires the Harmony response format. It fits on a single H100 GPU and supports fine-tuning, function calling, and tool use.

Source: HuggingFace — huggingface.co/unsloth/gpt-oss-120b-GGUF
Unknown
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
168.1k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerunsloth
ParametersUnknown
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads168.1k
Likes278
Last updated2025-08-25
Sourceunsloth/gpt-oss-120b-GGUF

What gpt-oss-120b-GGUF is

gpt-oss-120b is a Mixture-of-Experts language model trained on the Harmony response format. The base model is 120B parameters with 5.1B active parameters. This GGUF variant is quantized by Unsloth and includes optimizations beyond the original OpenAI release. It supports inference via Transformers, vLLM (with custom wheels), Ollama, and llama.cpp. The model requires manual Harmony format application when using model.generate() directly but is integrated into Transformers' chat template. Context length is unknown. Fine-tuning is stated as feasible on a single H100 node for the 120B variant.

Quickstart

Run gpt-oss-120b-GGUF locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="unsloth/gpt-oss-120b-GGUF")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Agentic reasoning tasks requiring transparency

Ideal for applications needing explicit chain-of-thought reasoning, function calling, and web/code execution. The full reasoning access facilitates debugging and auditing of multi-step agent decisions.

Production deployments on constrained infrastructure

Fits on a single H100 GPU with MXFP4 quantization; suitable for on-premises or cost-conscious cloud deployments where model size and inference speed matter.

Customized reasoning models via fine-tuning

Apache 2.0 license and stated fine-tuning support make this suitable for organizations needing to adapt reasoning behavior to domain-specific tasks without external model dependencies.

Running & fine-tuning it

**Estimate to verify:** 120B parameters with MXFP4 quantization in MoE layer fits on single H100 (80GB); full precision inference not practical on consumer hardware. Smaller gpt-oss-20b (21B parameters) can run within 16GB. Exact VRAM footprint depends on quantization level chosen (GGUF format supports multiple quantization schemes) and framework overhead. For CPU inference, llama.cpp via Ollama supports smaller quantizations on 32GB+ systems; production serving on GPU recommended.

Model card states full parameter fine-tuning is feasible on a single H100 node for 120B variant. LoRA/QLoRA support not explicitly mentioned but is standard in Transformers ecosystem. No guidance on optimal learning rates, batch sizes, or Harmony format preservation during fine-tuning provided. Practitioners should reference OpenAI's gpt-oss documentation and community examples.

When to avoid it — and what to weigh

  • You require guaranteed context window size — Context length is not disclosed in the model card. You must verify this against your application requirements before deployment.
  • You need inference on consumer hardware without quantization — 120B parameters in full precision exceeds typical consumer VRAM. Ollama and LM Studio offer alternatives, but only with quantization or smaller variants (gpt-oss-20b).
  • You cannot enforce Harmony response format in your pipeline — The model is trained exclusively on Harmony format and will not function correctly without it. This adds a mandatory dependency to your inference stack.
  • You require stable, out-of-the-box compatibility with all LLM frameworks — vLLM requires a custom wheel build from Unsloth's index; standard vLLM may not work. Integration maturity varies across frameworks.

License & commercial use

Apache License 2.0 (OSI-compliant, permissive). No copyleft restrictions, patent protection, or field-of-use clauses. License is clear and does not impose derivative work obligations.

Apache 2.0 is a permissive OSI license that explicitly permits commercial use, modification, and distribution. No patent risk or additional licensing required for commercial deployment. Model weights are open and redistributable. However, practitioners should verify compliance with any underlying training data licensing, OpenAI's terms of service for derivative works, and cloud provider terms if hosting commercially.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

No explicit security audit, adversarial robustness, or vulnerability disclosure details provided in the model card. As an open-weight model, the weights themselves are inspectable but the training process and safety measures are not fully transparent. Inference via external services (e.g., vLLM endpoints) introduces network/access control considerations. Prompt injection and hallucination risks are not addressed. Users should implement standard LLM safety practices: input validation, output filtering, rate limiting, and audit logging for production agentic workloads.

Alternatives to consider

gpt-oss-20b (OpenAI, same org)

Smaller variant (21B parameters, 3.6B active) with lower latency and consumer-hardware compatibility (16GB RAM). Trade-off: less reasoning capacity than 120B.

Meta Llama 3.1 (405B or 70B)

Also Apache 2.0; larger reasoning capacity and broader fine-tuning community. Trade-off: no native MoE optimization; requires more hardware for 405B variant.

Mistral Large or Mistral 7B

Apache 2.0 licensed; strong reasoning and agentic capabilities at different scales. Trade-off: different training approach; less emphasis on full chain-of-thought transparency.

Software development agency

Ship gpt-oss-120b-GGUF with senior software developers

Start with Unsloth's GGUF quantization for immediate inference, or fine-tune on your H100 for domain-specific reasoning. Verify Harmony format integration and context length for your use case. Contact Devco for production deployment architecture review.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

gpt-oss-120b-GGUF FAQ

Can I use gpt-oss-120b commercially without additional licensing?
Yes. Apache 2.0 is a permissive open-source license that permits commercial use, modification, and redistribution without additional licensing fees or restrictions. However, verify compliance with training data sources and your deployment environment's terms of service.
What GPU do I need to run gpt-oss-120b?
Single H100 (80GB) fits the model with MXFP4 quantization. Smaller consumer GPUs (e.g., 24GB L40S, RTX 4090) can run quantized GGUF variants with llama.cpp/Ollama, but inference speed will be reduced. For CPU inference, plan on high latency. Verify VRAM with your chosen quantization level before deploying.
Do I have to use the Harmony format?
Yes. The model was trained exclusively on Harmony response format and will not function correctly without it. If using Transformers' pipeline or chat template, Harmony is applied automatically. For raw model.generate(), apply it manually via Transformers' chat template or OpenAI's openai-harmony package.
Is the model actively maintained?
The base gpt-oss models are maintained by OpenAI. The GGUF quantization variants are maintained by Unsloth, with last update on 2025-08-25. Unsloth provides documentation and blog support. No explicit SLA is stated, but community adoption is active.

Custom software development services

Need help beyond evaluating gpt-oss-120b-GGUF? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and open-source llms integrations — and maintain them long-term.

Ready to Deploy gpt-oss-120b?

Start with Unsloth's GGUF quantization for immediate inference, or fine-tune on your H100 for domain-specific reasoning. Verify Harmony format integration and context length for your use case. Contact Devco for production deployment architecture review.