Lattice
Skip to content

Adapt open weights to my data

A fine-tuned model stack.

Fine-tuned model · 100,000 req/mo. At the default case — 100k requests a month, no further constraints — drawn from the 112 tools already in the index. Each pick is a tool whose own skip-when is stated, because a stack built on tools it does not know how to overcommit to is a guess.

This is the default case, stated as such. Describe your case to change it — the recommendation recomputes, and anything it assumes comes back in the answer rather than staying silent.

The picks

  • Training — Unsloth

    Strong fit

    LoRA fine-tuning on a single GPU, with memory and time roughly halved.

    Watch out: You need a training stack you can audit line by line.

    Consider Megatron-LM instead when tensor and pipeline parallel pretraining at cluster scale

  • Inference — Marlin

    Strong fit

    Near-GPU throughput from 4-bit weights, as a kernel inside another runtime.

    Watch out: You need a scheduler and a server as well as a kernel.

    Consider llama.cpp instead when the constraint is hardware, not throughput — no GPU, or an edge box

  • Prompts — Agenta

    Workable

    Open-source prompt management with a playground and versioning.

    Watch out: You need eval depth rather than prompt ergonomics.

    Consider DSPy instead when compiling prompts into optimised programs from labelled examples

  • Evals — Arize Phoenix

    Workable

    OpenTelemetry-native evaluation you can run yourself.

    Watch out: You want a vendor support contract behind it.

    Consider Braintrust instead when making evaluation, rather than tracing, the primary workflow

The costs and the risks

Estimated cost
$40–$220 per month — a heuristic band from query volume, not a vendor quote.
Confidence
87% — lower when constraints narrow the field.

Biggest risk at the default case: Fine-tuning teaches behaviour, not facts. If the answer changes monthly, it belongs in retrieval instead.

A starting point, not a prescription. The recommendation at the default case is what the engine would tell a stranger with your workload; the recommendation at *your* case is what it tells you once it knows the constraints. The Stack Builder holds both — the builder link preselects this workload and starts with your volumes from here.