# Tune an open model with LoRA on your own or rented GPU

> Tune an open model with LoRA on your own or rented GPU: 0 companies fully do it, 2 with a limit, 13 carry the tag. Every listing in Fine-tuning with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/fine-tuning.local-lora/
- Markdown: https://www.anchorterminal.com/niches/fine-tuning.local-lora/index.md (~1,000 tokens)
- Slim: https://www.anchorterminal.com/niches/fine-tuning.local-lora/index.min.md (~230 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/fine-tuning.local-lora/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 0 | 2 | 13 | 15 | 8 |

## What counts as done

A documented config or notebook for the chosen base model runs to completion on the stated GPU, the chat template and end-of-turn token are trained so the model stops, the docs state the minimum VRAM and supported accelerators, and the adapter is saved to disk.

Measured, for our own tests: a LoRA run on a 7-8B model over 1,000 examples finishes in under an hour on one 24 GB GPU.

How agents ask for it: fine-tune Llama on my chat logs; QLoRA Qwen on my GPU; train a small model for my app on my own machine

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Axolotl (Axolotl AI)**: YAML configs per model, chat_template with roles_to_train and train_on_eos, and NVIDIA (Ampere or newer) or AMD support are documented, but no VRAM table was found, and acc-08 saw a new model crash in the first forward pass and acc-09 an example whose end-of-turn token was not trained.
- **Unsloth (Unsloth)**: A VRAM table (8B: 6 GB QLoRA, 22 GB LoRA), NVIDIA, AMD, Intel and Mac support and get_chat_template with EOS mapping are documented, but acc-07 saw a script fail on a conversation format, acc-13 flat loss, and acc-47 no NPU backend.

- not this job: Amazon Bedrock model customisation (Amazon Web Services, Inc.): A hosted AWS service; training runs in Bedrock.
- not this job: Microsoft Foundry fine-tuning (Azure OpenAI) (Microsoft Corporation): A hosted service; training runs on Azure capacity.
- not this job: Fireworks AI Fine-tuning (Fireworks.ai, Inc.): Training runs on Fireworks GPUs, serverless or dedicated; there is no local trainer.
- not this job: Nebius Token Factory fine-tuning (Nebius B.V.): A hosted service; jobs run on Nebius GPUs.
- not this job: Tinker (Thinking Machines Labs, Inc.): The loop runs locally but training runs on Thinking Machines' GPUs; nothing trains on the user's own card.
- not this job: Together AI Fine-tuning (Together Computer, Inc.): A hosted service; training runs on Together's GPUs, not the user's.
- not this job: Vertex AI Gemini tuning (Google LLC): A hosted Google Cloud service; training runs on Google's accelerators.

## What stops agents finishing

- new model architecture crashes in the first forward pass
- end-of-turn token not trained so the model never stops
- accelerator (Ascend, Intel, some AMD) not supported

## Around it

Before: gpu-compute.find-and-provision, gpu-compute.training-run. After: fine-tuning.export-and-serve.

## Where this comes from

Weight 15 of 100 in Fine-tuning: 8 of 47 accounts: Unsloth and Axolotl issues (new-model crashes, templates, flat loss, accelerator support), Hacker News reports of quick Gemma tunes, and one forum thread on an on-device model

As data: https://www.anchorterminal.com/api/v1/use-cases/fine-tuning.local-lora.json, or the MCP tool get_use_case with id fine-tuning.local-lora. Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
