Skip to content
Language models

Lumi-Simy

A compact 4B agentic model with tool calling

3 min read

Lumi-Simy is a 3.97-billion-parameter language model by author joaohsim, fine-tuned to select external functions (tool calling), invoke multiple tools in parallel, build query chains with response reading, and decline tool calls when no suitable tool matches. It is built on the reasoning model Lumi-Think-4B, which originates from CEIA Lumi-4B and the hybrid NVIDIA Nemotron 3 Nano 4B architecture (combining Mamba-2, MLP, and attention mechanisms). Training was performed via LoRA SFT (fine-tuning a small set of additional weights with subsequent merging into the base model) while preserving the thinking mode and general skills in English and Portuguese. The model is conceived as a research and educational project for experimenting with agentic behavior on compact hybrid architectures.

Key features

  • Notable growth in multi-step tool calling: in the BFCL test, accuracy in multi_turn_base scenarios rose from 25.0 to 43.0, and in live_parallel_multiple from 41.7 to 70.8 compared to the base Lumi-Think-4B.
  • Parallel calls and declining unnecessary functions: the model can query multiple APIs in a single step and respond correctly without tool calls when the task does not require them.
  • Preservation of general knowledge: through the addition of self-distilled replay data during training, scores on general EN/PT tests remained at the level of the original Lumi-Think (average score 75.6 vs. 75.8).
  • Integration with a tool router: for large API catalogs, the author recommends pre-selecting the top 32 functions via the BAAI/bge-reranker-v2-m3 cross-encoder — this cuts the prompt from ~21k to ~5k tokens and boosts accuracy.

What you can do

  • Build a local AI agent for databases and dashboards: the model was tested on BI API and dashboard API scenarios (VAKRA benchmark) and can fetch data step-by-step from multiple sources.
  • Connect to services via OpenAI-compatible API: deploy via vLLM with qwen3_coder and nano_v3 parsers and use in chatbots where external functions need to be triggered in parallel (e.g., requesting weather or data for multiple objects at once).
  • Explore a reproducible fine-tuning pipeline: use the model card and weights as a reference for training compact Mamba hybrids on a single RTX 4090 class GPU (24 GB).

Example

hf download joaohsim/Lumi-Simy --local-dir lumi-simy
VLLM_USE_FLASHINFER_SAMPLER=0 vllm serve ./lumi-simy --trust-remote-code \
  --mamba-ssm-cache-dtype float32 --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser-plugin "$PWD/lumi-simy/nano_v3_reasoning_parser.py" \
  --reasoning-parser nano_v3 --default-chat-template-kwargs '{"enable_thinking": true}' \
  --max-model-len 32768 --max-num-seqs 16
Read moreDetails
VAKRA (5 new domains): 32.7% (Capability 1) and 51.7% (Capability 2; 60.3% with the bge-reranker-v2-m3 router vs. 16.5% for Lumi-Think-4B).
BFCL v3: outperforms Qwen3-4B across all four multi-turn categories, but trails in single and parallel calls.
Hardware and training: fine-tuned and tested on a single NVIDIA RTX 4090 (24 GB), training context of 20,480 tokens, 4,329 examples.
Limitations: educational release for non-commercial and academic use (VAKRA CC BY-NC-SA 4.0 license); not tested on web browsing or coding agents; reasoning in tool tasks is conducted primarily in English.
Discuss Telegram 0 28

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 7.4 GB · ready files9.3 GB
  • FP8 / INT8recommendedweights 3.7 GB5.6 GB
  • 4-bitweights 2.1 GB4.0 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 3080

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

Language models: with a 8192-token context.

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

Run it on Massed Compute

A ready script for the machine we advise above: paste it and the model downloads and starts.

  1. 1. Rent A30

    $0.26 an hour with the coupon. Pick the Ubuntu image when you create the machine.

    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.

    # joaohsim/Lumi-Simy — A30, FP8 / INT8
    # NVIDIA driver: CUDA 13 needs 580+. If older or missing, the current nvidia-open is installed (keep yours: KEEP_DRIVER=1)
    DRV=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader 2>/dev/null | head -1 | cut -d. -f1)
    case "$DRV" in ""|*[!0-9]*) DRV=0 ;; esac
    if [ "$DRV" -lt 580 ] && [ -z "${KEEP_DRIVER:-}" ]; then
      . /etc/os-release && \
      curl -fsSLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${VERSION_ID//./}/x86_64/cuda-keyring_1.1-1_all.deb" && \
      sudo dpkg -i cuda-keyring_1.1-1_all.deb && rm -f cuda-keyring_1.1-1_all.deb && \
      sudo apt-get update && \
      sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-open || \
      { echo "Could not install the NVIDIA driver: see the error above. To keep the current one, run with KEEP_DRIVER=1."; exit 1; }
      echo "NVIDIA driver updated. Reboot the machine (sudo reboot) and run this same script again."
      exit 0
    fi
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv venv ~/vllm --python 3.12 && source ~/vllm/bin/activate
    uv pip install vllm
    vllm serve joaohsim/Lumi-Simy --max-model-len 8192 --quantization fp8
    Add more LoRAs to Lumi-Simy
  3. 3. Use it

    The server listens on port 8000 and speaks the OpenAI API: http://localhost:8000/v1 on the machine. From your computer: ssh -L 8000:localhost:8000 user@machine-IP.

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

  1. NVIDIA-Nemotron-Nano-12B-v2-Basefamily root
  2. NVIDIA-Nemotron-Nano-9B-v2
  3. NVIDIA-Nemotron-3-Nano-4B-BF16
  4. lumi-4b
  5. Lumi-Simy

New posts about versions of NVIDIA-Nemotron-Nano-12B-v2-Base (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 3All on Hugging Face
Fine-tunes 2All on Hugging Face

About the model

joaohsim/Lumi-Simy
Author
joaohsim
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

280

Ask AI about this model

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

Comments