Open LLMs/unsloth

Open-Weight LLM · Private & Custom AI

gpt-oss-20b-unsloth-bnb-4bit

Apache 2.0 open-weight MoE reasoning model (20B params, 3.6B active) for self-hosted ops automation, agentic workflows, and fine-tuned custom AI—fits 16GB VRAM.

gpt-oss-20b is OpenAI's smaller open-weight model trained on the Harmony format, designed for local deployment, tool use, and reasoning tasks without copyleft or patent risk. For ops teams, it's a permissive alternative to closed APIs: run reasoning, chain-of-thought, and agentic logic in your own environment, fine-tune for domain work, and avoid data exfiltration.

20.9B
Parameters
apache-2.0
License (OSI/permissive)
Unknown
Context
66.8k
Downloads

Model facts

Developerunsloth
Parameters20.9B
Context windowUnknown
Licenseapache-2.0 — OSI/permissive
Tasktext-generation
GatedNo
Downloads66.8k
Likes38
Updated2025-08-08
Sourceunsloth/gpt-oss-20b-unsloth-bnb-4bit

Private deployment

Run gpt-oss-20b-unsloth-bnb-4bit in your own environment

Self-hosting architecture: download weights from HF Hub, run via Transformers, vLLM, Ollama, or LM Studio. Native MXFP4 + 4-bit quantization (this variant) fits ~12–16 GB VRAM on consumer/enterprise GPUs. Data stays in your datacenter; no API calls, no third-party inference logs. Requires Harmony format compliance and ops discipline around prompt/output handling.

Operational AI use cases

01

Internal Knowledge & Support Automation

Deploy as a private reasoning engine for customer support ticket triage, FAQ generation, and KB search augmentation. Chain-of-thought outputs help ops teams debug decisions; adjust reasoning level (low/medium/high) per ticket complexity. Fine-tune on internal docs to reduce hallucination and enforce domain rules.

02

Workflow & Process Agent

Use native function calling and tool integration to build internal ops agents: approve purchase requests, route docs, summarize meeting notes, orchestrate cross-team tasks. Agentic capabilities (web browsing, Python execution via extensions) let you wire it into your operational stack without API rate limits or vendor lock-in.

03

Finance & Compliance Document Processing

Structured output + reasoning capabilities support parsing invoices, contracts, and regulatory filings. Full chain-of-thought is auditable and not shown to end users—useful for finance teams needing explainable AI. Fine-tune on company-specific templates and compliance rules; keep sensitive data on-premises.

Custom AI

As a base for custom AI

Fully fine-tunable base for domain-specific applications: embed into a customer-facing product (SaaS, internal tool), adapt reasoning effort and prompt structure, and own the model weights. Apache 2.0 license permits commercial derivative products. Smaller footprint than gpt-oss-120b makes iteration and deployment faster for ops-focused AI.

In the operating system

Where it fits

Knowledge & Reasoning Layer in an AI OS: sits between your ops data (docs, tickets, transactions) and workflow agents. Provides fast, local reasoning (vs. cloud LLM calls), integrates with your function-calling layer, and serves as backbone for structured decision-making—e.g., classifier → gpt-oss-20b reasoning → approval workflow.

Data control & security

Self-hosting eliminates data exposure to third-party APIs. Sensitive customer records, internal processes, financial data stay on your infrastructure. No telemetry, no model training on your prompts. Compliance benefit: audit trails, data residency, GDPR/HIPAA alignment are your responsibility and under your control—not a model guarantee, but an architectural choice that open-weight enables.

Hardware footprint

Estimate (unsloth 4-bit quantized variant): ~10–14 GB VRAM (L40S, A100, H100, RTX 4090+). Full precision (BF16): ~42 GB. Inference latency: ~100–300ms per token on consumer/mid-range enterprise GPU depending on batch size and reasoning level. No official benchmark provided; test on target hardware.

Integration

Supports OpenAI-compatible API via Transformers Serve or vLLM (drop-in for existing integrations). Use Harmony format for chat templates; reference OpenAI docs. Function calling via schema-based definitions; tool integration tested with web browsing and Python execution. Ops tooling: wire into ticketing (Jira, Zendesk), knowledge bases (Confluence, custom DBs), and workflow engines (Zapier, n8n, custom orchestrators). Quantized variant (4-bit BitsandBytes) requires `torch`, `transformers`, `kernels`; verify GPU memory before deploy.

When it's not the right fit

  • You need sub-100ms latency at scale without GPU investment—20B + reasoning overhead slower than optimized smaller models or API calls.
  • Your ops task requires real-time web data or live external APIs; Harmony format and reasoning-centric design not optimized for rapid multi-tool orchestration.
  • You need support, SLAs, or enterprise indemnification—open-weight means community docs and GitHub issues only; no vendor backing.
  • Compliance requires model interpretability beyond chain-of-thought explanations; no formal security audit or cryptographic guarantees published.

Alternatives to consider

Meta Llama 3.1 70B

Larger, lower active-parameter overhead, broader fine-tuning history in ops. No native MoE or reasoning specialization; more general-purpose. Llama 2 license (commercial OK) but heavier inference cost.

Mistral 7B / Mistral Large (quantized)

Smaller footprint, faster inference, easier fine-tuning on consumer hardware. Less reasoning depth; better for classification and simple workflows. Apache 2.0 licensed.

Qwen2.5 14B

Competitive reasoning quality, similar size, strong function-calling support. Apache 2.0 licensed, active community. Less MoE-specific optimization but comparable ops-automation fit.

FAQ

Can I run this on my own servers without sending data to OpenAI or Hugging Face?

Yes. Download weights from HF Hub once, then run inference entirely on-premises using Transformers, vLLM, or Ollama. No data leaves your environment. Quantized 4-bit variant fits 12–16 GB VRAM on standard enterprise/consumer GPUs.

Am I allowed to use this in a commercial product or SaaS?

Yes. Apache 2.0 license permits commercial use, modification, and distribution. No copyleft, no patent claims. You can build customer-facing apps, charge for them, and use this as your inference backbone.

What's the difference between this quantized variant and the full model?

This is 4-bit quantized (BitsandBytes) by Unsloth; saves ~3–4x VRAM vs. BF16 full precision with minimal accuracy loss. Use this for resource-constrained deployments. Full precision available on HF Hub if you have GPU headroom.

Does the 'reasoning' feature mean this model is better at complex ops tasks?

Reasoning level (low/medium/high) adjusts output depth and chain-of-thought length. Higher reasoning = longer latency, more detailed reasoning, but not guaranteed correctness. Test on your ops workload; fine-tuning often matters more than reasoning level alone.

Build a Private Ops AI System with gpt-oss-20b

Deploy gpt-oss-20b on your infrastructure to automate support, workflows, and reasoning tasks without vendor lock-in. LLM.co helps you integrate open-weight models into a complete AI operating system—custom fine-tuning, agentic layers, and data residency built in. Let's design your private AI stack.