Skip to content
Other

Jalebi-Loop-35M

a compact model for decision-making and routing

2 min read

Created by Airavat-ai, this model features just 35M parameters and is designed not for text generation, but for efficient option selection, classification, and criteria scoring. It takes context and a list of options as input, outputting scores for each choice in a single pass, which eliminates hallucinations and prompt injections. The architecture supports dynamic computation depth via a repeating middle block (ranging from K=1 for maximum speed to K=4 for accuracy). This solution is ideal for developers tackling query routing, filtering, and rapid custom fine-tuning on a single video card.

Key features

  • Ultra-lightweight architecture: only 35M parameters, with VRAM consumption ranging from 0.3 to 0.6 GB.
  • Flexible depth (K): switch between K=1 (high speed) and K=4 (enhanced accuracy) during inference.
  • Three task types: choice processing, binary probabilities (noul), and criteria scoring (score).
  • Fine-tuning ready: simple customization for custom business processes and classifiers.

What you can do

  • Request routing: distribute incoming support tickets across departments (billing, fraud, support).
  • Automatic classification: filter short, structured requests and texts without the risk of generating unnecessary junk.
  • Option ranking: evaluate the relevance of pre-made options against specified criteria.
  • Fast inference: deploy the model in services with strict latency requirements and minimal VRAM consumption.

Example

import sys, numpy as np
sys.path[:0] = ["code", "code/eval", "code/model"]
from s1mb_adapter import load_checkpoint
from heldout_eval import build_scorer

model, tok, meta = load_checkpoint("Jalebi-Loop.pt", "tok24k.json", device="cpu")
scorer = build_scorer(model, tok, 512, "cpu", layout=meta["layout"])

record = {
    "case_id": "demo", "question_id": "demo", "task": "choice",
    "state": "Customer writes: my card was declined while travelling abroad",
    "question": "Where should this request go?",
    "options": ["billing team", "fraud team", "general support"],
    "option_ids": ["0", "1", "2"],
}
scores = scorer.score_records([record], 4)[0]["scores"][-1]
probs = np.exp(scores - scores.max()); probs /= probs.sum()
print(probs)
Read moreDetails
Metrics: on the S1MB English v1 benchmark (K=4), the average task metric is 9.87.
Hardware: on an NVIDIA L4 in K=1 mode (bf16), the model delivers up to 411–421 requests per second with peak memory usage up to 0.38 GB.
Limitations: the context window is limited to 512 tokens, and each option is capped at 40 tokens. The model is not a text generator and performs poorly on long reasoning tasks.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.1 GB · ready files0.6 GB
  • FP8 / INT8recommendedweights 0.1 GB0.6 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.

    # Airavat-ai/Jalebi-Loop-35M
    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 Airavat-ai/Jalebi-Loop-35M --local-dir ~/models/Jalebi-Loop-35M
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Jalebi-Loop-35M
  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 Jalebi-Loop-35M (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

Airavat-ai/Jalebi-Loop-35M
Author
Airavat-ai
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

Other

HMH-1-Cell

HMH-1-Cell
Other

Alcyone-W

Alcyone-W