Skip to content
Other

RetCast-JEPA

ultra-lightweight foundation models for time series forecasting

2 min readupdated

A family of tiny models ranging from 2.05M to 9.19M parameters, built from scratch on the RetNet2.0 architecture and the JEPA approach (Joint Embedding Predictive Architecture, where the model predicts future semantics in latent space rather than noisy raw values). It combines a delta-rule rewritable memory with embedding prediction, outperforming much heavier transformers across several tasks. The models were trained on diverse time series datasets in a zero-shot setup without seeing target test tasks beforehand. It is a great fit for researchers and engineers who need fast, accurate time series forecasting on low-end hardware.

Key features

  • Ultra-lightweight footprint: variants from 2M to 9M parameters take up just 8 to 35 MB in memory and require minimal compute.
  • High speed: runs inference significantly faster than bulky alternatives like Google TimesFM 200M.
  • RetNet2.0 + JEPA architecture: uses delta-rule memory and pattern prediction instead of step-by-step noisy data generation.
  • High accuracy: achieves low MSE and MAE scores on benchmarks while keeping VRAM usage minimal.

What you can do

  • Forecast time series: generate point and quantile forecasts for numerical sequences using ready-made model variants (`2m`, `5m`, `9m`).
  • Run recurrent rollouts: generate future states in latent space with constant $O(1)$ memory consumption.
  • Test architectures: explore model components (including RetNet and transformer ablations) for research experiments.
  • Integrate into pipelines: embed the models into PyTorch data analysis pipelines for GPU execution.

Example

import torch
from retcast import RetCastJEPA

# Загружаем модель RetCast-JEPA 2M с HuggingFace Hub
model = RetCastJEPA.from_pretrained("starkprateek1/retcast-jepa", variant="2m", device="cuda")

# Подготовка контекста (B, L)
context = torch.randn(4, 512, device="cuda")
with torch.no_grad():
    out = model(context, horizon_len=192, mode="chunkwise")

point = out["point"]          # Точечный прогноз (4, 192)
quantiles = out["quantiles"]  # Квантили прогноза
Read moreDetails
Benchmarks: RetCast-JEPA 2M achieves a Mean MSE of 0.3706 compared to 0.3878 for Google TimesFM 200M.
Hardware requirements: the 2M variant uses only 22.0 MB of VRAM during inference.
License: MIT.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.0 GB · ready files0.5 GB
  • FP8 / INT8recommendedweights 0.0 GB0.5 GB
  • 4-bitweights 0.0 GB0.5 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # starkprateek1/retcast-jepa
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download starkprateek1/retcast-jepa --local-dir ~/models/retcast-jepa
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to retcast-jepa
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of retcast-jepa (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

starkprateek1/retcast-jepa
Author
starkprateek1
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments