Skip to content
Language models

SautiLM-Jua-STEM

an offline Gemma 3-based STEM model for Android

2 min readupdated

SautiLM-Jua-STEM is an ultra-compact language model from LosingTek Research built to solve math, physics, chemistry, and biology problems completely offline. Based on Google Gemma 3 (270M-IT), it is compiled into a 272.36 MB .litertlm file that runs directly on smartphones via the LiteRT-LM runtime. Unlike heavy cloud models, it targets budget hardware and was trained on 750 verified instructions in French (60%) and English (40%), with Swahili support in its metadata. It is a solid fit for developers building offline educational apps and mobile tutors for low-end devices.

Key features

  • Minimal hardware requirements: runs on smartphones with Android 8.0+ and 2 GB of RAM on ARM Cortex-A53/A55 CPUs or a mobile GPU.
  • Ultra-compact footprint: the final sautilm-jua-stem.litertlm file weighs just 272.36 MB thanks to dynamic_wi8_afp32 quantization via litert-torch.
  • Targeted fine-tuning: adapted using LoRA (a lightweight adapter on top of the base model, r=32, alpha=64) with loss masking applied strictly to responses.
  • Modular ecosystem: part of the specialized SautiLM-Jua model family, where each module handles a dedicated subject domain.

What you can do

  • Integrate it into a Flutter mobile app using the flutter_edge_ai and flutter_edge_ai_litertlm plugins to build an offline school study assistant.
  • Deploy it as a local solver and explainer for math, physics, chemistry, and biology in regions without reliable internet access.
  • Benchmark the speed and generation quality of compact Gemma 3 models inside the LiteRT-LM runtime on entry-level Android devices.

Example

From the model card: drop the sautilm-jua-stem.litertlm file into a Flutter app via the flutter_edge_ai + flutter_edge_ai_litertlm stack to answer high-school and college questions across math, physics, chemistry, and biology in French and English, completely offline.

Read moreDetails
Architecture: Gemma3ForCausalLM (18 layers, hidden_size=640, context length cache_length=2048).
Training: 230 SFT steps (LoRA r=32, alpha=64) on 750 instructions, final loss — 0.0548.
Export & verification: litert-torch (dynamic_wi8_afp32, TFLite FlatBuffer + HF Tokenizer), LiteRT-LM Runtime check — 10/10.
Minimum hardware: Android 8.0+ (API 26+), 2 GB RAM, ARM Cortex-A53/A55 or mobile GPU.
Discuss Telegram 0 4

Will it run?

Working out the memory needs

This is a LoRA: it only runs on top of the base model unsloth/gemma-3-270m-it. The base model's size is still being collected; the estimate will appear within a day.

Massed ComputeOpen Massed Compute

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.

    # sautilm/SautiLM-Jua-STEM
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    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 unsloth/gemma-3-270m-it --local-dir ~/models/gemma-3-270m-it
    hf download sautilm/SautiLM-Jua-STEM --local-dir ~/models/SautiLM-Jua-STEM
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to gemma-3-270m-it
  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.

  • MTP (multi-token prediction)0.34.0from November 16, 2026new release

    1.5–2× generation speed

    If the model has MTP layers (DeepSeek-V3/R1, Qwen3-Next, GLM, MiMo and others), it predicts several tokens per step itself.

    How to enable: vLLM / SGLang: --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. Check the model card.

    Earlier generations (2)
    • EAGLE-3 — Speculative decoding with a small draft head. For models without their own MTP layers.
    • EAGLE-2 / Medusa — Earlier draft-head generations. Still work if that is all you find.
  • FP8 KV cache0.34.0from November 16, 2026new release

    2× more context in the same memory

    Keeps the attention cache in 8 bits, so long contexts fit where they did not.

    How to enable: vLLM: --kv-cache-dtype fp8; llama.cpp: -ctk q8_0 -ctv q8_0.

  • vLLM / SGLang0.34.0from November 16, 2026new release

    many times more tokens/s on a server

    Server engines: batching, FlashAttention, FP8 and speculative decoding out of the box.

    How to enable: vllm serve <model> — an OpenAI-compatible API.

    Earlier generations (1)
    • llama.cpp — GGUF models on any GPU, Mac or even CPU; some layers can stay in RAM.

Model family

Requires the base model

unsloth/gemma-3-270m-it

This model is a LoRA / adapter. Download the base model first, then load this one on top.

  1. gemma-3-270mfamily root
  2. gemma-3-270m-it
  3. gemma-3-270m-it
  4. SautiLM-Jua-STEM

New posts about versions of gemma-3-270m (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 7All on Hugging Face

About the model

sautilm/SautiLM-Jua-STEM
Author
sautilm
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

40

Ask AI about this model

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

Comments