/Tech3h ago

Ramp Labs releases PorTAL, a framework for transferring LoRA task adapters across different base LLMs without retraining

Story Overview

Ramp Labs introduced PorTAL as a hypernetwork setup that learns task latents once on a source model and then ports the adaptation to fresh base LLMs by training only a slim converter on limited calibration data, sidestepping full LoRA retraining each time a new model appears.

233852423850.2K

#684

Original post

elvis@omarsar0#684inTech

Highly-recommended read.

Even with all the research, fine-tuning is such an underexplored problem.

Based on what I've seen among the top AI-powered orgs, we are on the cusp of a fine-tuning revolution.

Agentic fine-tuning is going to dramatically change things in AI.

rahul@rahulgs

In many ways, finetuning or RLing a custom model is a bet against model progress and scaling. It's to choose to say "we don't think there's going to be a good enough base model for this task anytime soon, so we're not going to wait"

with oss release velocity these days, its a hard tradeoff

It's easy to end up on a custom model with an outdated base (Kimi 2.6 is only a few months old)

So we fixed it - PorTAL lets you swap base models quickly, allowing your learned task specific behaviors to port to new models as they come, no matter how fast

12:42 PM · Jul 1, 2026 · 6.3K Views

Cost Pressure

Half the data still hits the accuracy mark

Tests on Qwen3 variants recovered 94 to 98 percent of the usual LoRA accuracy lift while using roughly half the calibration examples, and the same pattern held when moving to an unseen Gemma-3 model.

Open Question

No public code or scaling checks yet

The announcement supplies no repo, weights, or results on models larger than 8B, instruction-style tasks, or third-party verification, so teams eyeing production use will need to wait for those details.

Sentiment

Positive users praise PorTAL for making fine-tuning viable across models and improving open source economics, while the negative reply expresses frustration over geopolitical neglect in AI access.

Pos

66.7%

Neg

33.3%

8 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Posts from X

Most Activity

VIEWS4.6KBOOKMARKS11LIKES65RETWEETS3REPLIES3

Florian Brand@xeophon

What’s your favorite neolab, anon?

Chinese DoorDash or American AliPay?

Ramp Labs@RampLabs

Introducing PorTAL: Portable Task Adapters for LLMs. A novel recipe to cheaply port fine-tuning between models. It matches per task LoRA accuracy at half the cost, lowering the switching overhead of adapting tasks across LLMs.

At Ramp, every new model release used to mean retraining our fine-tunes from scratch. PorTAL learns the task once, then efficiently refits it onto any new base model, even across model families.

2h4.6K6511

Johnny West@johnnywestt

@eglyman Eric this is pretty awesome

1h6811

elvis@omarsar0

@dpetrou

2h68

David Petrou@dpetrou

@omarsar0 What is "agentic" fine tuning?

3h49

Kartik@1kartikkabadi1

@eglyman this is a breakthrough for humanity. congrats big time! now every company can fine tune their own models and actually have it be justifiable

3h2982

oops@Joonzzy

@eglyman Ramp the frontier AI lab?

2h1532

kenneth@kennethnym

@xeophon american wechat

2h502

Steven Collard@stalmico

@eglyman so you retune once and carry it to the next Llama/Qwen drop?

2h204

Alex Mirran@alex_mirran

@eglyman Impressive

2h126

Xiangyi Li@xdotli

@xeophon Chinese Indeed vote +1

2h351

Vanar@Vanarchain

@omarsar0 Fine tuning still feels like one of the biggest untapped levers for production AI.

1h341

Mo@goesbymo

@eglyman as someone working with AI in a bunch of emerging markets, this may be what helps the world outside china/us/eu catch up.

1h89

WΞNDΞL@_wrbr

@xeophon lol... reading this and thinking: here on south america we are f*, americans and chineses dont give a f* to us. europeans are trying hard, at least.

2h37

Practically Perfect@practiklyperfct

@eglyman Interesting

1h34

Martin Höppner@MartinHoeppner

@eglyman @alexolegimas Well, that's timely!

18m13

Hira@Hiraweb3

@omarsar0 Agentic fine-tuning = models that learn to act, not just predict.

2h11

David Petrou@dpetrou

@omarsar0 Yes, but everything is done w agents now, so it felt to me that agentic fine tuning had a redundant word in it.

2h8

Eclipse 🌖@ECLresearch

@omarsar0 The alignment tax alone makes most current fine-tuning pipelines unsustainable at scale—curious how you see agentic fine-tuning sidestepping that overhead versus just shifting it to inference-time compute.

2h7

Winston B.@DoDataThings

Where I've seen it break is eval loops. Fine-tuning on agent trajectories only helps if your reward signal actually reflects task success and not just "did it call the right tool." I've watched get the tool-calling part right and then wonder why the model still fails multi-step tasks in prod. Scoring the whole trajectory can do it.

2h5

Peyton@PeytonAGI

@omarsar0 you can finetune into any modality

21m1