Starlings.Models.Lfm3
0.5.0
dotnet add package Starlings.Models.Lfm3 --version 0.5.0
NuGet\Install-Package Starlings.Models.Lfm3 -Version 0.5.0
<PackageReference Include="Starlings.Models.Lfm3" Version="0.5.0" />
<PackageVersion Include="Starlings.Models.Lfm3" Version="0.5.0" />
<PackageReference Include="Starlings.Models.Lfm3" />
paket add Starlings.Models.Lfm3 --version 0.5.0
#r "nuget: Starlings.Models.Lfm3, 0.5.0"
#:package Starlings.Models.Lfm3@0.5.0
#addin nuget:?package=Starlings.Models.Lfm3&version=0.5.0
#tool nuget:?package=Starlings.Models.Lfm3&version=0.5.0
Starlings.Models.Lfm3
Model license: LFM Open License v1.0, not MIT or Apache-2.0. The LFM3 weights, tokenizer and chat template are Liquid AI's. Commercial use is licensed only to a legal entity with annual revenue below USD 10 million (section 5). Above that threshold, LFM3 needs a separate commercial license from Liquid AI. Downloading the weights through this package means you accept the license. The full text is LICENSE.txt in this package and on the Hub. NuGet asks you to accept it when you install the package.
Text generation with LiquidAI/LFM3-1.7B on ONNX Runtime, through Liquid's own ONNX export (LiquidAI/LFM3-1.7B-ONNX at c680759afee1ce8310c90d6c621f4558200dee77). It includes the LFM3 chat template for text turns and a managed greedy decode loop over the export's step graph, with its convolution and attention caches. Prompts are tokenized by the owned tokenizer (Starlings.Tokenizers, profile lfm3-bpe-chat). The package ships the asset lock, not the weights: acquire them explicitly. The application adds the ONNX Runtime native package.
This library is experimental: admitted at Experimental (model-pipeline/admissions/lfm3.md, model request #217), with development evidence and no release gate.
Maturity
Maturity is per profile, generated from the registry manifests by model-pipeline catalog (do not edit this table by hand). Gates, evidence and capabilities per profile: the model catalog.
| Profile | Maturity | Platforms |
|---|---|---|
lfm3-1.7b/onnx-fp32-cpu |
experimental | linux-x64 cpu: experimental; linux-x64 cuda: untested; windows-x64 cpu: untested |
Install
dotnet add package Starlings.Models.Lfm3
dotnet add package Microsoft.ML.OnnxRuntime # CPU native runtime
What is checked
| Check | Result | Where |
|---|---|---|
Template text against transformers 5.21.0 apply_chat_template |
19/19 equal | Lfm3ChatTemplateTests (CI) |
| Prompt ids against the transformers tokenizer | 19/19 equal | Lfm3ParityTests (needs the snapshot) |
| Greedy ids and step-0 logits against the same fp32 graph on ONNX Runtime 1.30.0 in Python | 19/19 equal ids, logits max abs 0 | Lfm3ParityTests, evidence |
| Against the PyTorch checkpoint | not run | model-pipeline/reference/lfm3/run_lfm3.py reference |
Quick start
using Starlings.Generation;
using Starlings.Models.Lfm3;
// One-time setup: download the fp32 graph (7.2 GB) and the tokenizer into the asset store.
Lfm3ModelFiles files = await Lfm3ModelAssets.DownloadAsync(Lfm3Precision.Fp32);
// At startup: load offline.
using Lfm3ChatModel model = Lfm3ChatModel.Create(Lfm3ModelAssets.OpenLocal(Lfm3Precision.Fp32));
GenerationResult reply = model.Chat(
[ChatTurn.System("Answer in one short sentence."), ChatTurn.User("What is a heat pump?")],
new GenerationOptions { MaxNewTokens = 128 },
enableThinking: false);
Console.WriteLine(reply.Text);
With enableThinking: true (the template's default) the model writes its reasoning first and closes it with </think>. The completion text keeps that reasoning: splitting it off is the caller's job.
Capabilities
Lfm3Capabilities.All lists every capability with an explicit status. In short:
- Experimental: greedy generation on CPU (fp32), streaming (
GenerateStreaming), stop sequences and extra stop ids, chunked prefill, cancellation between calls, the text template withenableThinkingandpreserveThinking, asset acquisition. - Untested: the fp16 graph (
Lfm3Precision.Fp16) and CUDA, the intended RTX 3090 profile. The export's card notes that the CUDA provider has no fp32 GroupQueryAttention kernel, so CUDA needs the fp16 graph. Also untested: concurrent requests, contexts beyond the evidence prompts, and AOT. - Unsupported (rejected): sampling (the card's
temperature 0.1, top_k 50, repetition_penalty 1.1included), logprobs, guided decoding, tools and tool calls, image, video and audio input, prefix reuse, batches, the quantized graphs.
Lfm3Options.ContextLength defaults to 32,768 tokens. The 11 attention layers' caches grow by 44 KB per token in fp32. The checkpoint's 262,144 positions are accepted but untested.
License
The package carries two licenses; LICENSE.txt holds both, and the model's comes first.
- LFM Open License v1.0 (Liquid AI): the LFM3 weights and tokenizer, which the package downloads on request and does not contain, and
Lfm3ChatTemplate.cs, a C# port of the checkpoint'schat_template.jinjaand so a derivative work. Commercial use is licensed only below USD 10 million annual revenue (section 5(a) and 5(b)). Qualified non-profits are exempt for non-commercial or research use (5(c)). Full text. - MIT OR Apache-2.0 (StarlingHq): the rest of the package's code.
Attribution: THIRD-PARTY-NOTICES.md, also packed with the package.
| Product | Versions Compatible and additional computed target framework versions. |
|---|---|
| .NET | net10.0 is compatible. net10.0-android was computed. net10.0-browser was computed. net10.0-ios was computed. net10.0-maccatalyst was computed. net10.0-macos was computed. net10.0-tvos was computed. net10.0-windows was computed. |
-
net10.0
- Starlings.Acquisition (= 0.5.0)
- Starlings.Generation.Abstractions (= 0.5.0)
- Starlings.Models.Abstractions (= 0.5.0)
- Starlings.Runtime.Onnx (= 0.5.0)
- Starlings.Tokenizers (= 0.3.4)
NuGet packages
This package is not used by any NuGet packages.
GitHub repositories
This package is not used by any popular GitHub repositories.