Quick answer
Choose the runtime from the deployment job, not from a universal speed claim:
| Target | Starting runtime | Official artifact | Primary check |
|---|---|---|---|
| Local CPU or mixed CPU/GPU | llama.cpp | GGUF | Quantization quality, memory, prompt/decode speed |
| Apple Silicon | MLX | MLX | Unified memory, context growth, thermals |
| Direct Python and model access | Transformers | Native | Version, precision, device map, chat template |
| High-throughput GPU API | vLLM or SGLang | Native | Concurrency, tail latency, memory, API parity |
| Cross-platform edge | ONNX Runtime | ONNX | Execution provider, graph support, hardware acceleration |
Liquid AI publishes official native, GGUF, MLX, and ONNX repositories. Its model card documents Transformers, vLLM, llama.cpp, MLX, LM Studio, and SGLang paths. That is compatibility evidence, not a winner for your workload.
Local inference
Use GGUF with llama.cpp or a compatible local application when you need broad local hardware support and quantization choices. Use MLX when Apple Silicon is the deployment target and you want a package optimized for that framework. Pin the exact repository revision, file, runtime version, context setting, and launch flags.
Quantization reduces weight memory but can change quality. It does not remove KV-cache growth, runtime buffers, operating-system pressure, retrieval services, or tool-process memory.
GPU serving
The official model card provides vLLM and SGLang server examples for LiquidAI/LFM2.5-2.6B. Both can expose an OpenAI-compatible chat-completions surface. “Compatible” does not guarantee identical chat templates, tool-call serialization, error behavior, usage fields, batching, or cancellation.
Load-test representative prompts and tool traces at planned concurrency. Measure successful tasks, time to first token, decode rate, p50/p95 latency, peak memory, queueing, cancellation, retries, and response-schema parity.
ONNX edge path
Choose ONNX when the target deployment already depends on ONNX Runtime or a particular hardware execution provider. Validate the exact exported repository, model inputs and outputs, quantization, tool template, long-context support, and target accelerator. Do not extrapolate a cloud or Mac benchmark to an unrelated edge device.
Use one comparison protocol
- Pin one checkpoint and one prompt/task set.
- Use equivalent precision or disclose quantization differences.
- Warm up each runtime consistently.
- Test short, typical, and stress context lengths.
- Score output and tool-call correctness before comparing speed.
- Include installation, observability, crash recovery, upgrades, and rollback.
Use the local run guide to establish a baseline and the phone and CPU guide to size constrained devices.
Frequently asked questions
Which runtime should I use for LFM2.5-2.6B on a laptop?
Start with the official GGUF artifact and llama.cpp or a compatible app for general local use. On Apple Silicon, compare the official MLX artifact against GGUF on your real prompt and context lengths.
Which runtime should I use to serve LFM2.5-2.6B on a GPU?
The official model card documents vLLM and SGLang for high-throughput GPU serving with the native checkpoint. Benchmark the selected versions, concurrency, context, tool behavior, and operational controls.
Is ONNX the fastest LFM2.5-2.6B format?
No universal speed result follows from the format name. ONNX is the official cross-platform export; performance depends on the graph, execution provider, device, precision, context, and workload.
Official sources
- LiquidAI LFM2.5-2.6B inference matrix
- Liquid complete model and format library
- Liquid llama.cpp deployment guide
Source check: August 5, 2026. Verify current artifact repositories, supported runtime versions, licenses, accelerator support, and tool-call behavior.