LFM2.5-230M — Torq build (Synaptics SL2610-Series Torq NPU)

This repository provides compiled model files for LiquidAI's LFM2.5 230M text language model, ready to run on the Synaptics SL2610-series Torq NPU. The model can act as a skill-selection layer, taking natural-language instruction and decompose it into a sequence of tool calls.
Quick start guide:
- Buy a Machina kit: Get an SL2600 Machina kit delivered to you
- Torq Examples: Use Torq-examples LiquidAI/LFM2.5-230M scripts to download and deploy on your Machina kit

Model Overview
LFM2.5-230M is a general-purpose text-only model. It is a hybrid architecture that combines short convolutions with grouped-query attention.
The transformer runs on the Torq NPU in bf16; the token embeddings run on the host CPU.
Model Features
Contents
| File | Size | Role |
|---|---|---|
model.vmfb |
461 MB | monolithic build — decoder + LM head in one graph (logits output) |
body.vmfb |
327 MB | split build — decoder body only, emits hidden states (pairs with lm_head.vmfb) |
lm_head.vmfb |
134 MB | split build — standalone LM head (hidden → 65 536 logits) |
token_embeddings.npy |
134 MB | CPU embedding lookup table (bf16) |
config.json |
— | model configuration |
tokenizer.json, tokenizer_config.json |
— | tokenizer + tokenizer config |
onnx/model.onnx (+ model.onnx_data) |
~952 MB | reference ONNX export for non-Torq runtimes (e.g. onnxruntime) |
Monolithic vs. split
Two equivalent ways to run the model (same weights — body 327 MB + lm_head 134 MB ≈
the 461 MB monolithic build):
model.vmfb(monolithic): one graph that outputs logits directly. Simplest to run.body.vmfb+lm_head.vmfb(split): the decoder body emits hidden states and the LM head is applied only when sampling. Prefill tokens then skip the large[1024 → 65 536]LM-head projection, which lowers time-to-first-token — pick this when TTFT matters.
The onnx/ export is provided for reference / portability to other runtimes.
Model Details
- Architecture: LFM2 (
Lfm2ForCausalLM) — hybrid short-convolution + grouped-query attention. - Hidden size: 1024 · Layers: 14 · Attention heads: 16 (8 KV heads, GQA) · Intermediate size: 2560.
- Vocabulary: 65 536 · Context length: up to 128 k.
- Precision: bf16 on the NPU.
- Target: Synaptics SL2619, compiled with the Torq compiler.
Tested Platforms
Metrics
| Platform | Model / Stage | Environment | NPU Clock | Inference Time | Infer / s |
|---|---|---|---|---|---|
| SL2619 2GB | LFM2.5-230M | Torq v2.0.0 | 1 GHz | TBD | 7.1 |
Deployment
The models have been tested with the following environment.
- Torq Compiler: v2.0.0
- Torq Runtime: v2.0.0 included in Astra SDK release scarthgap_6.12_v2.4.0
Usage Tutorials / Example Apps
A usage example is provided in the Torq Examples / LiquidAI-LFM2.5-230M.
Check out the README for instructions.
Usage Notes:
- Place the model files in a directory and invoke the Torq LLM runner with either
model.vmfb(monolithic) orbody.vmfb+lm_head.vmfb(split, lower TTFT), alongsidetoken_embeddings.npy,config.json, andtokenizer.json.
License & attribution
This repository is a redistribution of a model created by Liquid AI, Inc., licensed under the LFM Open License v1.0. Copies of the license and the attribution notices are included alongside the model files:
- LICENSE — a verbatim copy of the LFM Open License v1.0.
- NOTICE — the copyright, patent, trademark, and attribution notices retained from the original Work (per Section 4(c) of the license).
Original model: LFM2.5-230M · Copyright © Liquid AI, Inc.
Learn More
- Synaptics AI Developer Zone: Get started with documentation, tutorials and resources for your Edge AI journey.
- Torq Compiler Documentation: Learn more about the Torq compiler based on MLIR and IREE.
- Synaptics Astra SDK: Learn more about the Yocto Project-based Linux software available for Astra SL processors.
- Astra Support Portal: Connect with our engineering team and community.
- Downloads last month
- 1,488