The whole "frontier models" conversation has gotten pretty weird. There's this unspoken assumption that only a handful of well-funded orgs can produce competitive models, and everyone else gets to be a consumer. The SovereignAI framing — an organization being able to build, deploy, and govern its own AI stack — gets thrown around a lot, but most discussions stop at vibes and don't give you anything actionable.
This technical report from tri-fair-lab takes a different angle. The core claim is that you can hit frontier performance through Continual Learning on open-weight base models rather than from-scratch pretraining. And they actually ship a model, Thomson 1.0, as proof.
What "Continual Learning" means here #
Most "sovereign AI" attempts I've seen fall into one of three buckets:
- Small-scale fine-tuning on a narrow domain
- Prompt engineering on a frozen API model
- Tool augmentation (
[RAG](/en/tags/rag/), function calling) bolted onto a closed model
None of these actually give you ownership of the weights. The Thomson approach is different — they take an existing open-weight model and apply a modern mid- + post-training stack to it, with two specific safeguards:
Plasticity preservation— so the model can still learn new things** Stability preservation**— so it doesn't catastrophically forget what it already knew
The design philosophy is "minimal high-impact interventions on the parameters" rather than throwing more data at it.
The π-shaped result #
The part that genuinely surprised me is the evaluation pattern. When you plot capability gains across domains, you don't get the usual narrow spike from fine-tuning. You get what they call a π-shaped curve: broad improvements across capabilities (including ones they didn't explicitly target) combined with near-elimination of the catastrophic forgetting problem that plagues narrow domain adaptation.
That second point matters a lot for the SovereignAI pitch. A law firm fine-tuning a model on case law doesn't want it to forget how to write Python. A tax consultancy doesn't want the model to lose its multilingual capabilities. Standard fine-tuning creates exactly this trade-off.
What's actually in Thomson 1.0 #
They position it as a general-purpose model with emphasis on high-stakes professional work — the domains that productivity research keeps flagging as AI-susceptible: agentic tasks, legal, tax, safety, multilingual, and large-scale Deep Research. Evaluations show it competing with recent frontier releases on a wide range of benchmarks.
The compute and personnel budget disclosure is the other interesting part. They explicitly state the results were achieved with resources "substantially lower than commonly thought," which is the real thesis: the gap between "Big Lab frontier model" and "mid-size org with a focused continual learning pipeline" might be much smaller than the discourse suggests.
Why this matters beyond the paper #
If continual learning on open weights is actually viable at frontier-ish quality, it changes a lot of procurement conversations. In-house data stays in-house. Tool infrastructure gets to be custom. Values alignment isn't outsourced to a model's RLHF team you'll never meet. It's not going to replace a 100k-GPU pretraining run for cutting-edge reasoning, but for the 90% of enterprise use cases that need a good model that the org actually controls, the calculus looks different now. Worth reading the full report if you're working on AI strategy at any org that's tired of being a pure API customer.
[Computer Anthology: The Benchmark That Grows With AI 21d ago](/en/news/4950/)
[Next LLM-generated spam is basically a DDoS attack on open-source →](/en/news/7685/)