# Take the tuned weights out and serve them elsewhere

> Take the tuned weights out and serve them elsewhere: 0 companies fully do it, 6 with a limit, 10 carry the tag. Every listing in Fine-tuning with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/fine-tuning.export-and-serve/
- Markdown: https://www.anchorterminal.com/niches/fine-tuning.export-and-serve/index.md (~1,100 tokens)
- Slim: https://www.anchorterminal.com/niches/fine-tuning.export-and-serve/index.min.md (~280 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/fine-tuning.export-and-serve/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 | 6 | 10 | 12 | 6 |

## What counts as done

The adapter and a merged checkpoint can be downloaded in a documented format (safetensors, PEFT, GGUF) under a stated licence, and the docs show the export loading in at least one named runtime, including for vision and MoE models.

How agents ask for it: export my fine-tune to GGUF; load my LoRA in vLLM; run my tuned model in Ollama; download the weights

## 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)**: axolotl merge-lora then axolotl export writes f16 and llama.cpp quantised GGUF files, with preflight checks, but it needs a separately built llama.cpp and refuses pre-quantised checkpoints; licence is the base model's.
- **Fireworks AI Fine-tuning (Fireworks.ai, Inc.)**: firectl model download writes a model to a local directory, but the reference states no format or licence and no page read shows the export loading in a named runtime.
- **Nebius Token Factory fine-tuning (Nebius B.V.)**: Every checkpoint's result_files can be downloaded, the models page states each base model's licence, and a guide merges MoE LoRA weights; but no page read names the file format or shows it loading in vLLM or llama.cpp.
- **Tinker (Thinking Machines Labs, Inc.)**: save_weights_external and weights.download fetch trained weights, but the README names no file format, licence or runtime that loads the export.
- **Together AI Fine-tuning (Together Computer, Inc.)**: Download merged, adapter or raw output as a .tar.zst by job ID or step, but the example archive shows pytorch_model.bin, and no page read shows the export loading in a named runtime such as vLLM or llama.cpp.
- **Unsloth (Unsloth)**: save_pretrained_merged writes merged 16-bit or LoRA files, GGUF export and a vLLM guide are documented, but acc-14 and acc-15 hit GGUF conversion failures, acc-16 and acc-17 vLLM refusing tuned models, and acc-18 undocumented MoE LoRA packing.

- not this job: Amazon Bedrock model customisation (Amazon Web Services, Inc.): Custom models are served inside Bedrock through on-demand deployments or Provisioned Throughput; no weight download appears in the customisation pages read.
- not this job: Microsoft Foundry fine-tuning (Azure OpenAI) (Microsoft Corporation): The pages read deploy tuned models to Serverless, Managed Compute or Fireworks on Foundry; job result_files carry a metrics CSV, and no weight download is offered.
- not this job: Vertex AI Gemini tuning (Google LLC): Tuned Gemini models exist as endpoints, with adapters stored by Google; no weight download is offered for Gemini tunes.

## What stops agents finishing

- GGUF conversion fails inside the vendor's own image
- vLLM cannot load the saved vision or MoE checkpoint
- hosted service keeps the weights

## Around it

Before: fine-tuning.local-lora, fine-tuning.managed-job. After: local-ai.private-offline-run, gpu-compute.serverless-model-endpoint, open-weights.local-coding-agent.

## Where this comes from

Weight 12 of 100 in Fine-tuning: 6 of 47 accounts: five Unsloth issues on GGUF conversion, vLLM loading of vision tunes and MoE-LoRA layouts, plus a Together integration that serves and registers the result

As data: https://www.anchorterminal.com/api/v1/use-cases/fine-tuning.export-and-serve.json, or the MCP tool get_use_case with id fine-tuning.export-and-serve. 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
