Platform · Model Studio

One control plane for the entire model lifecycle.

Train, slim, benchmark, govern, and release private models from one workspace. Every build is a registry entry — base, method, eval, channel, compliance state, and a signed passport — so the model you ship is the model you can defend.

Enter the studioSee the lifecycle

90/100

Best hard-benchmark readiness (32B v1)

83/100

Served-tier readiness (7B v3 · production)

2,800

Verified training examples · 0 customer rows

0

Customer data rows in any corpus

The lifecycle

Four stages, one registry

A model moves from a training run to a signed release without leaving the system of record. Each stage hands the next a versioned artifact and the evidence behind it.

System of record

Every model is a registry entry

A model is not a file on someone's laptop. In the studio it is a typed record with the lineage that makes it auditable — what it was trained from, how, on what data, how it scored, where it sits, who approved it, and a signature you can re-check later.

The same record drives the studio UI, the release gate, and the procurement evidence pack. There is no separate spreadsheet to keep in sync — the registry is the source of truth.

Promotion through dev → candidate → approved → production is a state change on the record, gated by eval verdict and a reviewer who is not the trainer. Nothing reaches a channel it has not earned.

Example passport · SprintLoop-7B v3

BaseQwen2.5-Coder-7B-Instruct
MethodLoRA · MLX (Apple Silicon)
Datasetsprintloop-corpus · 2,800 · no customer data
Eval83/100 hard suite · 0 leakage
Channeldev → candidate → production
Reviewreviewer ≠ trainer (separation of duties)
Passportseals on artifact + hash · re-verifiable

Why a control plane

The studio replaces the gap between a good run and a defensible release

Most teams can produce a model. Far fewer can prove, months later, exactly what it was and why it was allowed to ship. The studio closes that gap by construction.

Governed data in, no customer rows

Datasets are registered, PII-scanned, and marked for customer-data status before a run can use them. The SprintLoop corpus is 2,800 verified examples with zero customer rows.

Hard probes, not vibes

Each version is scored on a behavioral, safety, and leakage suite. The 32B candidate reads 90/100 with zero leakage; the served 7B reads 83/100, pass-with-one-warning — both visible, neither hidden.

Separation of duties

The approver is never the trainer. A release that has not cleared a reviewer who is independent of the build cannot enter the production channel.

A passport that seals into proof

On release, a passport is sealed with an Ed25519 signature over its lineage and its stored artifact hash — so a regulator or customer can re-check it without trusting your dashboard. Until that artifact is stored, the registry shows the build as passport-ready, not signed.

Technique catalog

The methods the studio exposes

Live methods run on SprintLoop models today. Available methods are wired and ready to select. Planned methods are on the roadmap. Status is a property of the technique, not marketing — live reads in the accent.

Fine-tune

LoRA / QLoRAadapter tuning on MLX — SprintLoop-7B/32B todayLive
DoRAweight-decomposed LoRA — higher fidelity, ~LoRA costAvailable
GaLoregradient low-rank — full-rank quality at LoRA memoryAvailable

Slim

GGUF / AWQ / GPTQ / FP8quantize to lean — SprintLoop-7B · Edge (Q4)Live

Distill

On-policy distillationshrink a teacher into a leaner studentAvailable

Align

DPO / GRPOpreference + reasoning RL — sharpen refusal behaviorAvailable
RLAIF / self-rewardingAI feedback loop — scale alignment without labelersPlanned

Prune

Wanda / SparseGPTone-shot pruning — drop weights, keep qualityAvailable

Merge

TIES / DARE / SLERP mergecombine adapters/variants without retrainingAvailable

Continual

Autoresearch loopmeasure → modify → keep/discard → repeat — running on SprintLoopLive

See the lifecycle running on real models

The studio holds a registry of builds today, from a production 7B adapter to a compliance model in its first governed draft. Open the command center and follow one through every stage.

Enter the studioStart with training