Origin Swap: move a Qwen or DeepSeek fine-tune onto US open weights and prove parity

I found nothing that does exactly this: take a Qwen or DeepSeek fine-tune, re-run it on several US open bases, and produce a parity and cost report for buyers. Every part already exists, though. Tinker, OpenPipe, Oumi and Predibase handle multi-base fine-tuning, eval tools handle side-by-side comparison, and the bases themselves are well covered (Nemotron, Meta's Muse, Thinking Machines' Inkling, and Nvidia's reported Poolside deal). The demand signal is real. Salesforce publicly argued for Nemotron because its data provenance is clear and Qwen's is not, and state and government bans on Chinese models put pressure on startups selling to regulated buyers. A new entrant's best angle is the procurement-ready artifact, not the training. That artifact would be a reproducible, signed parity report: the same SFT data and evals, pinned weight revisions, the cost to serve, and a lineage scan from Cisco's open-source kit proving the new base really is non-Chinese. That last point matters because some Nemotron variants are themselves built on Qwen weights. The weaknesses are that the training step is easy for platforms to commoditise, each customer probably needs it once rather than on an ongoing basis, and the evidence that Chinese models are behind on capability is mixed. The second idea, the standalone disclosure page, is weak. Trust centers like Vanta and SafeBase can add the same fields easily, and lineage fingerprinting means buyers will expect verified evidence rather than self-attestation. It works better as an output of the swap tool or as a plugin for those trust centers than as its own product.

• Thinking Machines Tinker

  A hosted fine-tuning and RL API that charges per token. You could run the same dataset on several bases with it, including Qwen and their own US model, Inkling, but it has no built-in origin-swap or parity report. It is more likely the backend this CLI would call than a direct rival.

• OpenPipe

  A fine-tuning platform that A/B tests fine-tuned candidates against your production model and supports trying different foundation models. It is the closest workflow match, but nothing I found is built around origin or US-base parity.

• Oumi

  An open-source platform for training and evaluating models across many bases, with about $10M raised. A team could script a multi-base LoRA sweep with it, but it is not packaged as a migration product.

• Predibase LoRA Land / LoRAX

  Showed same-config LoRA fine-tunes on 10 base models with low cost per run, and its LoRAX server hosts many adapters on one GPU. This is the method this CLI would automate, but it is not offered as a provenance-driven product.

• Cisco Model Provenance Kit

  A free open-source CLI that fingerprints weights to check model lineage, and it found that some Nemotron models are built on Qwen weights. That is a direct risk for the disclosure-page idea: buyers can now verify lineage instead of trusting a self-written page. It also means a model labelled American may still have Chinese lineage.

• Vanta / SafeBase trust centers

  These already host subprocessor and AI sections and automate security questionnaires. A base-model disclosure page fits naturally as a feature they could add, so a standalone generator has a thin moat.

• Google Cloud model migration eval guide

  A notebook guide for comparing models systematically, including fine-tuned against base models. It is free how-to material, not a product, and shows that a basic parity eval is well-understood DIY work.

Checked 2026-10-17