veri

Frontier intelligence,
specialized for your use case.

Rapidly iterate on creating and deploying your own model, trained on your data.

Training Jobs

qwen-math-grpo-livejob_grpo_math_
running

Qwen/Qwen3-4BGRPO7/10/2026

2× MI300Xrunpod

Cost$42.1800
Duration120.0 min
Progress820/1200
Loss0.1837
View metrics in W&B
llama-support-sftjob_sft_suppor
queued

meta-llama/Meta-Llama-3-8B-InstructSFT7/10/2026

1× A100-80GBrunpod

Train models on your data.
Built for teams that need specialized models,
Veri handle all the infrastructure complexity.

01

Finetune

Bring your own dataset and kick off a run in a few lines. Veri provisions the exact GPU config, handles checkpointing, and streams logs end to end.

02

Deploy

Promote a model to a live endpoint with one call. Scale up and down on demand and ship your agent to users in minutes.

Train effortlessly.

Bring your own dataset and kick off a run in a few lines. Veri provisions the exact GPU config, handles checkpointing, and streams logs end to end.

Training guide

Deploy peacefully.
Save massive.

Specializing a small open model on your task and it can match frontier quality at a fraction of the inference cost.

We provide an OpenAI compatible endpoint for deployed models, so you could seamlessly switch.

Deployment guide

Scale up without
the complexity

Run forward passes, rollouts, and training environments across one GPU or hundreds.
We handle the infrastructure for you, so you can easily focus on the model.

SFT w/ Unsloth

A practical first specialization run with LoRA adapters and a lightweight single-GPU path.

Go to docs →

Slime GLM-4.7-Flash

See work with Slime, GLM-4.7-Flash, on a Veri-provisioned 8xH100 cluster.

Go to docs →

Verl GSM8K PPO

A walkthrough through GSM8K PPO quickstart for reasoning eval loops using Verl.

Go to docs →

Meant to be
easy.

Run an experiment in a few lines of code with the Veri SDK, then compare results before they reach production.

SDK
from veri_sdk import training_job@training_job(    base_model="Qwen/Qwen2.5-0.5B-Instruct",    dataset="hf:openai/gsm8k",    gpu_type="L4-24GB",    gpu_count=1,)def reward(completions, answer, **kwargs):    return [1.0 if a in str(c) else 0.0            for c, a in zip(completions, answer)]
Read the docs →

Maximize Capability.

Iterate faster.

Get Started