§ 00 Training · What it is
Training means teaching a network new behaviour by moving its weights — not running a finished model. On Tenstorrent hardware that work is not CUDA. Their Custom Training track teaches tt-train (ttml), the autograd layer that gives TT-NN a backward pass. AGICY’s planned Vasilikos campus would host Galaxy-class systems for that kind of job. The campus is pre-construction. GPU Bridge is partner GPUs you can rent now — not campus Galaxy.
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. We do not claim to be first to offer Tenstorrent.
§ 01 Paths · Fine-tune · scratch
Tenstorrent treats these as two different jobs. Fine-tune unless you have a specific reason not to.
Inference uses a tool someone else already built: load weights, feed inputs, read outputs. Training builds or reshapes the tool — slower, experimental, and only worth it when you need behaviour the base model does not have. That is the distinction their Understanding Custom Training lesson opens with.
Fine-tuning starts from a pre-trained model and specialises it. From-scratch training starts from random weights. The latter is how you study every piece of a network, or research a new design; it is rarely the enterprise default. Fine-tuning on this site covers PEFT / LoRA on the planned stack. This page is the wider training map — including when you would not fine-tune.
| Fine-tuning | From scratch | |
|---|---|---|
| Starting point | A finished open-weight model | Random weights; the model has not learned yet |
| What you spend | Hours to days on a focused example set | Far more data and wall-clock time — usually a research or lab choice |
| When Tenstorrent recommends it | Specialise a model for a task or domain; this is the default | New architecture, full control, or teaching yourself every piece |
| AGICY path (design) | Planned Galaxy jobs after COD; PEFT detail on /solutions/fine-tuning | Phase 2 evaluation (e.g. wafer-scale) — not Phase 1 live capacity |
§ 02 Stack · Three TT ways
There is not one training stack. Picking the wrong one is the usual source of confusion — including on this site, previously.
Tenstorrent’s autograd training framework lives inside the TT-Metalium source tree. Python bindings are called ttml. TT-NN ops compute a forward result only — they do not record how to differentiate it. ttml wraps those ops with a matching backward pass and an on-device AdamW optimiser so a real loop (forward, loss, backward, update) can run on Tenstorrent silicon. ttml is built from source; it is not a pip wheel. This is what the VS Code toolkit Custom Training lessons teach.
A different, actively maintained project of training recipes on the TT-Forge / TT-XLA compiler path — not a settings layer on top of tt-train. The two do not share code or config. Reach for it if you already work in that compiler world. The Custom Training track does not teach it.
Most teams have trained this way: a DataLoader, loss.backward(), an Adam-family optimiser. ttml is written so that four-step loop still feels familiar; the hardware and the library that know how to run backward change. AGICY GPU Bridge is this familiar GPU path on partner machines — not Tenstorrent Galaxy, not Vasilikos.
Tenstorrent also has a separate “build every piece by hand” arc (tokenizer through the training loop) before handing off to ttml on Blackhole hardware. That is not this page, and it is not AGI Academy. Primary source: docs.tenstorrent.com — Understanding Custom Training.
§ 03 Loop · How a run works
High level only — enough to brief a buyer, not a copy of Tenstorrent’s lesson graphs.
Turn raw text into a training set. Tenstorrent’s lessons use prompt/response pairs in a simple line-oriented format. Quality of examples matters more than raw count.
Load existing weights (fine-tune) or start from random numbers (from scratch). Fine-tune unless you have a specific reason not to.
Take a batch, predict, score the error, compute gradients, update weights, repeat. That is the same four steps you would write in PyTorch — on Tenstorrent, ttml is what makes backward and the optimiser run on-device.
Spot-check generations, save weights and optimiser state, then serve the result. Tenstorrent points production serving at vLLM after a training run — that is inference, not another training job.
Multi-chip work on Tenstorrent is documented as data-parallel training: split a batch across chips, average gradients, keep every device’s weights the same. Their lessons start everyone on a single chip (n150, p100, or p300c) and cover n300, T3000, and Galaxy meshes later. Combining data-parallel and tensor-parallel axes is an advanced tt-train config — not something AGICY operates today. See their Multi-Device Training lesson. We do not republish their chip-count speedups as AGICY benchmarks.
Concept lessons need no hardware. Hands-on runs need a TT-Metalium tree built with tt-train enabled. QuietBox images ship TT-NN and vLLM; they do not pre-install that source tree. None of that is a live AGICY campus service.
§ 04 AGICY · Planned vs now
Distinct paths. Planned or design-target unless marked live. Training on Tenstorrent silicon is not the same product as Academy workshops.
tt-train / ttml on Tenstorrent after COD
Design-target home for custom training and then serving those weights on Tenstorrent Galaxy (Wormhole / Blackhole class systems). Vasilikos is pre-construction: not a live hall, not live MW, not executed offtake. OEM supply under diligence — not a Tenstorrent endorsement, and not a claim that AGICY is first to offer this silicon.
hardware/tenstorrent →PyTorch-style jobs on partner GPUs
EU-region partner GPUs you can rent now, until campus COD. Separate SKU from SRA. Not Vasilikos, not campus Galaxy, not tt-train on AGICY silicon. Use this when you need a familiar GPU training or fine-tune job this month.
pricing →Try inference — not a training cluster
OpenAI-compatible gateway plus a public playground with daily limits. You can try supported models today. That is inference, not a weight-update job on campus or on GPU Bridge.
gateway/playground →Workshops and course hours
Published modules (including a Custom Training course) and vertical workshops. Education and partner enablement — not a running tt-train job, not campus Galaxy, and not a substitute for Tenstorrent’s own VS Code lessons if you have their hardware.
academy →Large full-pass / continued train
Wafer-scale optionality for jobs that need a much larger training surface than a Galaxy PEFT run. Quote-only watchlist — not Phase 1 CapEx, not a live training floor.
hardware/cerebras →Future campus entitlement
Sovereign Resource Allocation is how buyers reserve planned Cyprus capacity. It is not a live training SLA and not GPU Bridge hours.
sra →§ 05 Compare · Jobs, not hype
Same word, three products. This table replaces the old live-fleet vs hyperscaler scorecard.
| Weight-update training | Inference | Academy | |
|---|---|---|---|
| What the job is | Move weights (fine-tune or from scratch) | Run a finished model | Teach people — not a cluster job |
| Tenstorrent software | tt-train/ttml, or tt-blacksmith, or PyTorch on GPU | TT-NN and vLLM are the usual serve path | VS Code toolkit lessons (vendor docs) |
| AGICY today | GPU Bridge: partner GPUs, not campus Galaxy | Copperway playground + gateway | Academy curriculum for SRA partners |
| AGICY after COD (plan) | Galaxy-class Tenstorrent at Vasilikos | Serve trained or open weights on campus | Same Academy — still not the cluster |
We do not claim a 5× cost cut, a 671B-parameter max, Slurm/Kubernetes as a live AGICY training fabric, or that EU RISC-V training is available “for the first time.” Those were marketing lines, not Tenstorrent documentation and not a campus you can book this week.
§ 06 Questions · FAQ
§ 07 Start · Now · later
GPU Bridge is live partner compute. Campus Tenstorrent training is a planned product — apply for SRA if you need future EU-sovereign weight-update jobs.