Your AI model knows too much. Delete what it doesn't need.

AI models hoard redundant knowledge — noise, biases, overfitted patterns. Remap Studios surgically removes unnecessary neurons, shrinks the architecture, and retrains a leaner model in a fraction of the compute.

0+

Models unlearned

across research labs and companies

0.0%

Avg compute reduction

without accuracy loss

0.0M

GPU hours saved

equivalent to $6.3M in compute

0.0%

Avg accuracy retained

after selective unlearning

0+

Supported models

HuggingFace compatible

0.0K

Developers

using Remap Studios

See the difference unlearning makes

Select a model, drag the slider, and watch the metrics update in real time. Every number reflects real benchmarks from our test suite.

Before Unlearning

13.5 GB

7B parameters

After Unlearning

9.3 GB

50% smaller

BEFORE13.5 GB
AFTER9.3 GB
Accuracy retained: 99.2%
Unlearning Intensity
None50%Maximum
MODEL SIZE

9.3 GB

50% reduction

TRAINING COST

$2,436

42% savings

RETRAINING TIME

3d

vs 6 weeks original

INFERENCE SPEED

2.3×

faster inference

Cost Breakdown (per training cycle)
GPU hours$2940 → $1705
Storage$630 → $365
Engineering time$630 → $365
Total saved$1,764 (42%)

Explore how a neural network works

Click any layer in the diagram to see what it does, how many parameters it has, and whether it's a primary target for unlearning.

Interactive Architecture — Click any layerLLaMA-style Transformer · 6.7B params
FORWARD PASSEmbedding Layer154MSelf-Attention Q38MSelf-Attention K38MSelf-Attention V38MAttention Output38MLayer Norm 18KFFN Gate92MFFN Up92MFFN Down92MLayer Norm 28KLM Head154M
Layer Breakdown
Total Parameters6.7B

See the network. Then shrink it.

Hover over nodes to inspect them. Toggle between the original bloated network and the unlearned, optimized version.

All nodes erased. Toggle to rebuild.

Active nodes
Erased nodes

How unlearning works — from start to finish

Click through each stage to understand the complete pipeline. Every step has real metrics from production runs.

STEP 01

Upload Model

Load any HuggingFace-compatible model in .safetensors or .pt format. The system analyzes architecture, counts parameters, and maps every tensor.

1 / 5
model.safetensors⬆ Drop file hereor click to browseOpen Model
Formatssafetensors, .pt, .bin
Max size70B parameters
Parse time< 5 seconds

The unlearning pipeline

Four steps from a bloated, expensive model to a lean, fast one. Every step is automated — you just point at your model.

01

Analyze the Model

Map every node, weight, and connection in the neural network. Identify which neurons are essential and which are redundant, noisy, or overfitted.

02

Identify Redundancy

Our algorithm detects dead neurons, overlapping representations, spurious correlations, and knowledge that no longer serves the model's purpose.

03

Erase Unnecessary Nodes

Surgically remove the identified nodes and connections. The network architecture shrinks while preserving core functionality.

04

Retrain the Lean Model

The shrunken model retrains in a fraction of the original time. Less nodes = less compute = same accuracy, dramatically lower cost.

Erase Nodes Yourself

This is a simplified neural network. Click hidden-layer nodes to erase them and watch the compute savings in real time. Use Auto-Erase for bulk removal.

NODES

0 / 0

COMPUTE SAVED

0%

INPUTHIDDEN 1HIDDEN 2HIDDEN 3OUTPUT

Click any hidden-layer node to erase it. Use Auto-Erase to bulk-remove redundant nodes.

Works with the models you already use

Verified benchmarks across popular open-source models. Any HuggingFace-compatible causal LM is supported.

LLaMA 2

Meta

Verified

7B / 13B / 70B

REDUCE

38–52%

ACCURACY

99.1%

SPEED

2.3×

Mistral 7B

Mistral AI

Verified

7B

REDUCE

41–48%

ACCURACY

98.8%

SPEED

2.1×

CodeLlama

Meta

Verified

7B / 13B / 34B

REDUCE

35–55%

ACCURACY

99.3%

SPEED

2.5×

Falcon

TII

Verified

7B / 40B

REDUCE

32–47%

ACCURACY

98.5%

SPEED

1.9×

Phi-2

Microsoft

Verified

2.7B

REDUCE

44–60%

ACCURACY

99.0%

SPEED

2.8×

Gemma

Google

Verified

2B / 7B

REDUCE

36–50%

ACCURACY

98.9%

SPEED

2.2×

Qwen 2

Alibaba

Beta

0.5B–72B

REDUCE

30–58%

ACCURACY

98.7%

SPEED

2.0×

DeepSeek

DeepSeek

Beta

7B / 67B

REDUCE

33–49%

ACCURACY

98.6%

SPEED

2.1×

Calculate your ROI

Input your model size, training frequency, and GPU costs. See exactly how much you'll save annually.

Model Size
7B
Training Runs / Month
4
Avg GPU Hours / Run
48h
GPU Cost / Hour
$2.50
Expected Reduction
45%
ANNUAL SAVINGS

$5,443

95% reduction in compute costs

PER-TRAIN SAVINGS

$113

per training run

GPU HOURS SAVED/YR

2,177

hours of GPU time

MONTHLY SAVINGS

$454

per month

CO₂ REDUCTION

1089 kg

CO₂ equivalent saved

ANNUAL COST BEFORE$5,760
ANNUAL COST AFTER$317

Based on 4 training runs/month with 48 GPU hours each

Estimate compute savings

Fine-grained calculator for compute units, training steps, and cost per cycle.

100 nodes
20300
35%
5%80%
3 epochs
120
100 epochs
10300
$2.50
COMPUTE REDUCTION

98%

By removing 35% of nodes and retraining for just 3 epochs instead of 100.

BEFORE

$250.00

Original training cost

AFTER

$4.88

Retraining cost

YOU SAVE

$245.13

per training cycle · Model shrunk to 65 nodes

ORIGINAL10000 units
AFTER UNLEARNING195 units

Ready to make your model leaner?

Remap Studios is in early access. Download the desktop app or join the waitlist for API access.

SOC 2 Compliant
Self-hosted option
No data leaves your infra