§ 00 Fine-tuning · What it is
Fine-tuning takes a model that already knows language and teaches it your job — your tone, your terms, your task format. Cheaper than training from scratch. More durable than prompt tricks alone. AGICY’s planned path is EU-sovereign PEFT (LoRA, QLoRA) on RISC-V hardware; Copperway playground is what you can try today.
Honesty: Vasilikos is a design-target campus, not a live hyperscale facility, not live MW, and not executed offtake. Not crowdfunding. Not a token sale.
§ 01 Why it matters · Your data
A general model is a strong starting point. It is not yet your model.
Off-the-shelf models are trained on public internet-scale text. They can draft, summarise, and code — but they do not know your house style, your product codes, or the way your counsel writes a memo. Fine-tuning is how you bake that in, using examples you already have, without paying to pre-train a foundation model.
For EU buyers the location of that job matters. Fine-tuning copies patterns from your examples into weights (or into a LoRA adapter). If that run happens on a US-parented cloud, you are putting sensitive material through another jurisdiction. AGICY’s plan is to keep those jobs on a Cyprus (EU) campus design — when the facility reaches COD. Until then, evaluate the stack, reserve SRA capacity, and use Copperway playground to try models.
Fine-tuning is not always the first move. If your facts change every week, retrieval-augmented generation (RAG) is usually cheaper and easier to audit. Fine-tune when the model must internalise a stable skill, tone, or format that prompting and RAG do not hold.
§ 02 Compare · vs training · RAG
Same family of work. Different cost, risk, and when you should pick each.
| Fine-tuning | Pre-training | RAG | |
|---|---|---|---|
| Starting point | A finished open-weight model | Random weights; the model has not learned yet | A finished model plus your documents at query time |
| What changes | Some or all of the model weights (or a small adapter) | Every weight, from scratch | Nothing in the model. Retrieved text is added to the prompt |
| Data you need | A smaller, labelled set that shows the job you want | Huge corpora — usually impractical for one enterprise | A living knowledge base you can update without retraining |
| Best when | You need the model to speak your domain, tone, or task format | You are building a new foundation model | Facts change often, or you need citations from source files |
| Cost & time | Hours to days with PEFT (LoRA / QLoRA); longer for a full pass | Weeks to months and a dedicated training cluster | Ongoing retrieval cost; no weight-update job |
| AGICY path (design) | Planned Galaxy PEFT jobs; CS-3 eval for large full passes | Phase 2 wafer-scale / Helios evaluation — not Phase 1 live capacity | Planned sovereign RAG pipeline — see /solutions/rag |
PEFT is not a fourth column — it is how most fine-tuning is done now. Instead of moving every weight, you train a small adapter (LoRA / QLoRA). Full fine-tuning is the heavier option when adapters fall short. Training on this site means weight-update jobs (fine-tune or from scratch) on the planned Tenstorrent fleet after COD — tt-train / ttml, still pre-COD, not a live cluster. GPU Bridge is partner GPUs, not campus Galaxy.
§ 03 How it works · LoRA · PEFT
High level only — enough to choose a method, not a research paper.
You begin with a pre-trained open-weight model (Llama, Mistral, DeepSeek, Qwen, Gemma, and similar). You prepare a focused dataset: prompts and the answers you actually want, or ranked pairs for preference methods. The run then updates weights so the model’s next answers look more like those examples.
Teams iterate: learning rate, how many passes, adapter size, evaluation on held-out examples. The goal is a model that generalises to new questions in the same domain — not one that only memorises the training file.
Parameter-efficient fine-tuning (PEFT) freezes most of the original model and updates only a small slice. That is cheaper, faster, and less likely to wipe out what the base model already knows. LoRA and QLoRA are the PEFT methods most buyers start with.
LoRA (Low-Rank Adaptation) trains a thin adapter instead of rewriting every weight. You can swap adapters for different jobs without storing a full copy of the model. QLoRA goes further by using a compressed (quantised) base, which cuts memory. Typical PEFT jobs use far less compute than a full fine-tune — the saving depends on the model and data, not a fixed 90% promise.
Every weight can move. Use this when adapters are not enough — for example a large domain shift. It needs more memory, more time, and more care so the model does not forget its general skills. On AGICY’s roadmap this is the kind of job Cerebras CS-3 is evaluated for, not a live Vasilikos cluster today.
Instruction tuning shows the model example questions and good answers so it follows requests instead of just completing sentences. RLHF and DPO go further: they steer the model toward preferred behaviour (helpful, on-policy, less rude) using human or ranked feedback. These are post-training steps on top of a base or PEFT run — not a substitute for clean data.
§ 04 Hardware · Planned stack
Distinct paths a buyer can evaluate. Planned / design-target unless marked live.
PEFT / LoRA and serving the result
Design-target Phase 1 silicon: air-cooled Galaxy servers on a compiler-first RISC-V stack (TT-Forge path). This is the planned home for domain adapters and then inference of those weights — not a live 1,801-server hall. OEM supply under diligence; not a Tenstorrent endorsement.
hardware/tenstorrent →Large full fine-tunes / continued train
Wafer-scale optionality for jobs that need a much larger training surface than Galaxy PEFT. Different job than Phase 1 inference. Watchlist / quote-only — not Phase 1 CapEx, not a live training floor.
hardware/cerebras →Inference next to regulated ledgers
Trusted AI beside core banking, claims, or citizen records that should not leave the machine room. Telum II + Spyre. Not a substitute for Galaxy fine-tune clusters. US vendor; quote-only; no IBM endorsement.
hardware/ibm-z →Facility, not a chip SKU
Galaxy’s GDDR6 / Ethernet design is meant to run on standard air cooling — no mandatory liquid loop for the Phase 1 inference plan. The campus is pre-construction: not live MW, not executed offtake, not a hyperscale hall you can book this week.
data-centers →Try models now; not campus fine-tune
Copperway is the EU-oriented OpenAI-compatible gateway. The public playground lets you try supported models with daily limits. It is not production fine-tuning on Vasilikos silicon and not a token sale.
gateway/playground →§ 05 Choose · Which stack
A practical map. If two rows apply, do the cheaper one first.
| If you need… | Evaluate | Note |
|---|---|---|
| Teach tone, format, or a stable domain (legal style, product SKUs) | PEFT / LoRA on planned Galaxy | Start with adapters. Add RAG if facts still change weekly. |
| Documents update constantly and you need citations | RAG first | Fine-tune later only if the model still mishandles your format. |
| Adapters are not enough; you need a large weight update | Full fine-tune — CS-3 evaluation path | Treat as a Phase 2 quote, not a Phase 1 live job. |
| Model must sit beside the ledger / core transactions | IBM Z / LinuxONE watchlist | Inference next to data — not campus GPU training. |
| You want to try a model in the browser today | Copperway playground | Demo surface with limits. Reserve SRA for future campus capacity. |
§ 06 Research · Related briefs
Deeper briefs already on this site — hardware, sovereignty, and open-weight deploy.
Jim Keller's TT-Forge thesis, Blackhole T6 cores, and why predictable sovereign inference workloads favour compile-time planning over runtime GPU traffic control.
HARDWARE PARTNERGeneral availability April 2026. Jim Keller's vision and why AGICY's planned 1,801-server fleet is a design-target RISC-V commitment — pre-construction, not live.
HARDWARE WATCHLISTDated CS-3 / WSE-3 deep-dive. CS-4 announced 18 Aug 2026. Living page /hardware/cerebras. Phase 2 eval — not Phase 1 CapEx.
HARDWARE WATCHLISTGA 12 Aug 2026: Telum II + Spyre for AI beside the ledger. Phase 2+ watchlist — not Galaxy CapEx. US vendor; quote-only.
INFRASTRUCTURE ECONOMICSGalaxy server economics: GDDR6 memory, on-die Ethernet scale-out, and air-cooled TCO for EU sovereign campuses.
DEPLOYMENT GUIDEFrom foundation model to sovereign IP asset. Deploy, fine-tune, and own open-weight models on planned RISC-V infrastructure.
COMPLIANCEGDPR-compliant AI hosting requires EU-only data residency, Art. 28 DPAs, and no third-country transfers.
SOVEREIGNTYThe CLOUD Act problem: why 'sovereign' labels from US hyperscalers don't protect EU data, and what true sovereignty requires.
§ 07 Questions · FAQ
§ 08 Start · Try · reserve
Playground is live. Campus fine-tuning is a planned product — apply for SRA if you need future EU-sovereign adaptation jobs.