LFM2.5 2.6B at 4-bit generates a median of 113.0 tokens per second on the AMD Strix Halo 128GB, the hardware it is most run on, from one plain run (community reports, as of 25 Oct 2026).
As of 25 Oct 2026, LFM2.5 2.6B at 4-bit on the hardware it is most run on, with the median of plain runs (one device, one request, no speculative decoding, the whole model in its memory):
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
coding
User reports Liquid's latest model (LFM2.5) at 25 t/s sustained decode and 28 t/s sustained prefill at long contexts on an Intel 12th gen i5 with 24GB of mismatched DDR4 system RAM.
The run was CPU only; no engine, quant or context length is stated.
User calls it solid for spec driven development and asks why it gets no attention as an edge-deployment breakthrough.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is shown only when the source reports it; generation speed cannot tell us time to first token. Check the full setup before comparing.
agentictool-use
User reports LFM2.5-24B-A2B at ~95 t/s sustained on an AMD Radeon Pro W7900.
Setup is llama.cpp with ROCm, Q8 quant, full 128k context, taking 30GB VRAM.
User swapped it in for Ornith 1.5-35B in Pi Agent workflows and says it holds up well for mechanical and long-running recursive tasks; insight testing in council mode is pending.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is shown only when the source reports it; generation speed cannot tell us time to first token. Check the full setup before comparing.
agentictool-usemultilinguallong-context
User reports LFM2.5-24B-A2B at 184 tok/s on an Apple M5 Max.
Setup is GGUF weights with a 131,072-token context length.
The figure is the vendor's on-device claim; the same release is quoted at up to 21K output tok/s on a single NVIDIA B300 GPU. The model is a 24B-parameter MoE with 2.45B active parameters, 64 experts per block with top-4 routing, and 36 trillion training tokens.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
User reports LFM2.5-8B-A1B at 563.12 t/s generation and 30,680.66 t/s prompt processing on an RTX 5090.
Setup is llama.cpp llama-bench (build b9653) with Q8_0 weights, all layers on GPU (ngl=99), single request, no speculative decoding.
The dataset card also lists llama-bench rows for gpt-oss-20b, Nemotron-3-Nano-30B-A3B, Qwen3.6-35B-A3B, Qwen3-Coder-Next, Gemma 4 12B, Qwen3.6-27B, Gemma 4 31B, and Ling-3.0-flash, plus rows for Qwen3.6-27B-MTP, Qwen3.6-27B Q6_K, gpt-oss-120B, Mistral Small 4 119B, and several quant ladders; some runs used vLLM and SGLang.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
agentictool-use
User compares LFM2.5 2.6B and MiniCPM5 2B on small agentic tool-use tasks on an M1 Air.
LFM2.5 2.6B runs at 200 t/s prompt and 22 t/s generation with llama.cpp, QAD-Q4_0 GGUF, 32k context, using about 2.5 GB of RAM. MiniCPM5 2B runs at 162 t/s prompt and 16 t/s generation with a DSpark draft model, Q8_0 GGUF, 32k context, q5_1 KV cache, using about 3.8 GB of RAM.
User calls LFM2.5 the clear winner: faster, less RAM despite more parameters, and better aligned for English conversation. MiniCPM5 is noted as often responding in Chinese and hallucinating less, but LFM2.5 tends to hallucinate as conversations get longer.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
User reports LFM2.5-230M at 192.34 t/s generation and 1611.26 t/s prefill on a Xiaomi 14T Pro phone.
Setup is llama.cpp (build 96ffdc41c) running CPU only with 8 threads, the Q4_0 GGUF, and a q8_0 KV cache, measured with llama-bench at 256 prompt and 256 generation tokens over 10 runs.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is shown only when the source reports it; generation speed cannot tell us time to first token. Check the full setup before comparing.
agentic
User reports a custom inference engine running on a OnePlus 13, pure CPU.
The engine is built from scratch, is 450kb, and supports other model architectures.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
tool-usesummarization
User benchmarks 20 small models on an RTX 4050 6GB.
Models tested include Granite 4.1 3B, Gemma-4-agentic-e2b, Nemotron-3-Nano-4B and LFM2.5-8B-A1B.
The user selects LFM2.5-1.2B-Instruct as a cheap always-on model.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is shown only when the source reports it; generation speed cannot tell us time to first token. Check the full setup before comparing.
tool-useagentic
User reports Liquid AI's LFM2.5-2.6B at 113 tok/s on a Ryzen AI Max+ 395, a CPU speed from the vendor's own benchmarks.
Setup is the official Q4_K_M GGUF of around 1.67 GB, a 2.69B parameter model with 128K context, using under 2.5 GB memory in the vendor's tests.
The vendor also reports 30 tok/s on a phone and 220 tok/s on an M5 Max, and the user notes independent results are needed. Benchmarks are ToolSandbox 77.83, IFBench 59.17, BFCLv4 56.88 and LiveCodeBench 59.41, and the model card does not recommend it for agentic coding.
Individual observations from this page, not expected speeds or a ranking. Select a model to open its report, with the source and offloading conditions. Reported context may be a limit rather than actual input length.
The hardware LFM2.5 2.6B at 4-bit is most run on: AMD Strix Halo 128GB (113.0 t/s median, one run); M1 8GB (22.0 t/s median, one run). Medians of plain runs, as of 25 Oct 2026.
Which inference engines do people use for LFM2.5?
Reports of LFM2.5 name llama.cpp (5 reports), LM Studio (1 report), as of 25 Oct 2026.
Speed by hardware on LFM2.5 2.6B at 4-bit
The size and quant level the most devices have plain runs of: one device, one request, no speculative decoding, and the whole model in the device's memory. Every speed in this table is the same model.