Catalogue

43 models you can fine-tune.

Which models can I fine-tune on Onrup?

43 open-weight base models across 12 families, from 350M to 117B parameters. Each one lists the memory it needs for half-precision and four-bit training, which objectives it supports, its licence, and the cheapest GPU class that will run it.

The catalogue is curated rather than open. That is a real constraint and worth checking before you plan a migration: if your base model is not here, this is not the right platform for it.

Memory figures are admission thresholds — the point below which a run will not be scheduled because it would fail. They are deliberately conservative and are not a claim about the theoretical minimum.

SmolLM

Hugging Face

Small models trained on an unusually large token budget for their size, aimed at running where a GPU is not guaranteed to exist.

Model	Params	Type	LoRA	QLoRA	Objectives	From
SmolLM2 1.7B	1.7B	Dense	6 GB	4 GB	SFT, DPO	$0.09/hr
SmolLM3 3B	3B	Dense	10 GB	6 GB	SFT, DPO	$0.09/hr

Qwen3

Alibaba

The broadest size ladder in the catalogue, from 0.6B to a 30B mixture-of-experts, with permissive licensing across the whole range.

Model	Params	Type	LoRA	QLoRA	Objectives	From
Qwen3 0.6B	600M	Dense	4 GB	3 GB	SFT, DPO, GRPO	$0.09/hr
Qwen3 1.7B	1.7B	Dense	6 GB	4 GB	SFT, DPO, GRPO	$0.09/hr
Qwen3 4B	4B	Dense	12 GB	8 GB	SFT, DPO, GRPO	$0.09/hr
Qwen3 8B	8B	Dense	18 GB	14 GB	SFT, DPO, GRPO	$0.09/hr
Qwen3 14B	14B	Dense	28 GB	20 GB	SFT, DPO, GRPO	$0.09/hr
Qwen3 30B-A3B	30B / 3B active	Mixture of experts	64 GB	36 GB	SFT	$0.42/hr
Qwen3 32B	32B	Dense	64 GB	48 GB	SFT	$0.42/hr

Llama

Model	Params	Type	LoRA	QLoRA	Objectives	From
Llama 3.2 1B	1B	Dense	4 GB	3 GB	SFT, DPO	$0.09/hr
Llama 3.2 3B	3B	Dense	10 GB	6 GB	SFT, DPO	$0.09/hr
Llama 3.1 8B	8B	Dense	18 GB	14 GB	SFT, DPO, GRPO	$0.09/hr
Llama 3.3 70B	70B	Dense	140 GB	48 GB	SFT	$0.42/hr
Llama 4 Scout 109B	109B / 17B active	Mixture of experts	220 GB	80 GB	SFT	$1.68/hr

Mistral

Mistral AI

Apache-2.0 throughout, with a long-context 12B and a 24B that was positioned as best-in-class at its size on release.

Model	Params	Type	LoRA	QLoRA	Objectives	From
Mistral 7B v0.3	7B	Dense	16 GB	12 GB	SFT, DPO, GRPO	$0.09/hr
Mistral Nemo 12B	12B	Dense	24 GB	16 GB	SFT, DPO	$0.09/hr
Mistral Small 3 24B	24B	Dense	48 GB	36 GB	SFT, DPO	$0.42/hr

Phi

Microsoft

Data-quality-first models under a plain MIT licence, which consistently benchmark above their parameter count.

Model	Params	Type	LoRA	QLoRA	Objectives	From
Phi-4 Mini 3.8B	3.8B	Dense	12 GB	8 GB	SFT, DPO	$0.09/hr
Phi-4 14B	14B	Dense	28 GB	20 GB	SFT, DPO	$0.09/hr

Gemma

Google

A four-rung ladder from 1B to 27B. Supervised fine-tuning only in this catalogue; preference objectives are not offered on Gemma bases.

Model	Params	Type	LoRA	QLoRA	Objectives	From
Gemma 3 1B	1B	Dense	4 GB	3 GB	SFT	$0.09/hr
Gemma 3 4B	4B	Dense	12 GB	8 GB	SFT	$0.09/hr
Gemma 3 12B	12B	Dense	24 GB	18 GB	SFT	$0.09/hr
Gemma 3 27B	27B	Dense	56 GB	36 GB	SFT	$0.42/hr

DeepSeek-R1 distills

DeepSeek

Reasoning-distilled versions of existing 7B and 8B bases. The reasoning behaviour is already present before you start, which shortens the path for maths and code work.

Model	Params	Type	LoRA	QLoRA	Objectives	From
DeepSeek-R1 Distill Qwen 7B	7B	Dense	16 GB	12 GB	SFT, DPO	$0.09/hr
DeepSeek-R1 Distill Llama 8B	8B	Dense	18 GB	14 GB	SFT, DPO	$0.09/hr

LFM2

Liquid AI

Hybrid architectures combining gated convolutions with attention, built for on-device and agentic workloads. Licence is Apache-2.0-derived with a commercial revenue threshold.

Model	Params	Type	LoRA	QLoRA	Objectives	From
LFM2 350M	350M	Hybrid	3 GB	2 GB	SFT	$0.09/hr
LFM2 700M	700M	Hybrid	4 GB	3 GB	SFT	$0.09/hr
LFM2 1.2B	1.2B	Hybrid	5 GB	4 GB	SFT	$0.09/hr
LFM2 2.6B	2.6B	Hybrid	8 GB	6 GB	SFT	$0.09/hr
LFM2 8B-A1B	8.3B / 1.5B active	Mixture of experts	18 GB	12 GB	SFT	$0.09/hr
LFM2 24B-A2B	24B / 2.3B active	Mixture of experts	48 GB	28 GB	SFT	$0.42/hr

Granite

IBM

Enterprise-oriented models under Apache-2.0, tuned toward retrieval and tool use rather than open-ended chat.

Model	Params	Type	LoRA	QLoRA	Objectives	From
Granite 3.0 2B	2B	Dense	7 GB	5 GB	SFT, DPO	$0.09/hr
Granite 3.0 8B	8B	Dense	18 GB	14 GB	SFT, DPO	$0.09/hr
Granite 4.1 8B	8B	Dense	18 GB	14 GB	SFT, DPO	$0.09/hr

gpt-oss

OpenAI

OpenAI's open-weight mixture-of-experts models, Apache-2.0, with a small active parameter count relative to their total size.

Model	Params	Type	LoRA	QLoRA	Objectives	From
gpt-oss 20B	20.9B / 3.6B active	Mixture of experts	40 GB	24 GB	SFT	$0.17/hr
gpt-oss 120B	116.8B / 5.1B active	Mixture of experts	220 GB	80 GB	SFT	$1.68/hr

Falcon 3

TII

A clean four-rung ladder from 1B to 10B under the TII Falcon licence, with consistent behaviour across sizes.

Model	Params	Type	LoRA	QLoRA	Objectives	From
Falcon 3 1B	1B	Dense	4 GB	3 GB	SFT, DPO	$0.09/hr
Falcon 3 3B	3B	Dense	10 GB	6 GB	SFT, DPO	$0.09/hr
Falcon 3 7B	7B	Dense	16 GB	12 GB	SFT, DPO	$0.09/hr
Falcon 3 10B	10B	Dense	22 GB	16 GB	SFT, DPO	$0.09/hr

OLMo

Allen Institute for AI

Fully open models: training data, code and every intermediate checkpoint are published, not just the final weights. The right choice when provenance has to be auditable.

Model	Params	Type	LoRA	QLoRA	Objectives	From
OLMo 2 7B	7B	Dense	16 GB	12 GB	SFT, DPO	$0.09/hr
OLMo 3 7B	7B	Dense	16 GB	12 GB	SFT, DPO, GRPO	$0.09/hr
OLMo 3 32B	32B	Dense	64 GB	48 GB	SFT	$0.42/hr

Reading the objective column

SFT — Supervised fine-tuning. The default. Available on every model in the catalogue.
DPO — Preference tuning (DPO). Needs memory for a frozen reference model alongside, so roughly double the supervised figure.
GRPO — Reinforcement learning (GRPO). Requires a programmatic reward function; generates and scores several candidates per step.

How to choose between them →

Last verified 6 August 2026.

Start with the free tier

A magic link creates your account, your tenant and your first API key. No card until you ask for compute.

Get your API key Read the docs