Fourbitfourbit
Fine-tuning infrastructure

Four bits to a deployed model.
Record · train · eval · deploy.

Fourbit is the managed fine-tuning platform for LLMs, VLMs, VLAs, and diffusion models. We own all four stages so you don't have to — your data goes in, a working deployed model comes out.

H100 · H200 · B200 capacityLoRA · QLoRA · Full FT · DPO · ORPOPer-second billing
NewFourbit for robotics: managed Pi-0 fine-tuning + on-hardware deploy with our SO-101 pilot kit.
Learn more →
API & CLI

Launch a fine-tune in four lines.

Use the CLI for one-offs, the SDK for pipelines, or the dashboard when you want to click. Same primitives, same job IDs across all three.

  • Resumes from the latest checkpoint on preemption.
  • Streams loss, grad-norm, and sample generations in real time.
  • Promotes a checkpoint to a private endpoint with one call.
~ fourbit train
$ fourbit train \
  --base qwen3-8b \
  --data s3://acme/support-tickets.jsonl \
  --recipe lora \
  --budget 40usd

→ job ftjob_8r2hQv queued · 4×H100 · eta 38m
→ stream: https://fourbit.ai/r/ftjob_8r2hQv
The four bits

Four bits to a deployed model.

Record, train, eval, deploy. Every fine-tune in fourbit moves through these four stages — and we own all of them so you don't have to.

01

Record

Bring your data. We handle the rest.

Upload teleop episodes, JSONL, image sets, or point us at a bucket. We tokenize, shard, and pack — and tell you up-front if your data is the bottleneck.

02

Train

Curated recipes per architecture. Per-second billing.

LoRA, QLoRA, full FT, DPO. The right hardware lands automatically. Resumes from checkpoint on preemption. Stream loss live.

03

Eval

Real numbers, not vibes.

Sim eval, task-specific benchmarks, side-by-side with the base model. The scorecard your boss can read — before anything ships to production.

04

Deploy

Tuned weights are not a product. A working endpoint is.

One click to a private inference endpoint, or pull weights for your stack. 30-day on-call engineer included with every pilot.

Why fourbit

The boring parts, handled.

Recipes that just work

Curated training configs per architecture. No more chasing flash-attn versions or DeepSpeed YAML.

Eval that means something

Run task-specific evals on every checkpoint. Compare runs side-by-side, promote winners.

Bring your hardware

Run on our pooled H100/H200 capacity, or attach your own cluster. Same dashboard either way.

Pay for what you train

Per-second GPU billing, transparent pricing. Cap your spend before you queue.

Train your first model today.

Free credits to get your first run on the dashboard. No card needed.