LocalFit
Submit a benchmark

LFM2.5-1.2B-Instruct

LiquidAI/LFM2.5-1.2B-Instruct

Released2026-01-05Size1.2B tensors · denseRoleAssistantLicenselfm1.0Context125K config max

Where it runs 8K context · ≥ 5 t/s floor

SetupBest planMemoryGenerate
8 GB GPU + 32 GB RAMFP16llama.cpp CUDAFull GPUFits*~68–99 t/s Estimated
12 GB GPU + 32 GB RAMFP16llama.cpp CUDAFull GPUFits*~115–123 t/s Estimated
16 GB GPU + 64 GB RAMFP16llama.cpp CUDAFull GPUFits*~238–291 t/s Estimated
24 GB GPU + 64 GB RAMFP16llama.cpp CUDAFull GPUFits*~241–302 t/s Estimated
32 GB GPU + 128 GB RAMFP16llama.cpp CUDAFull GPUFits*~378–604 t/s Estimated
2× 24 GB GPUs + 128 GB RAMFP16llama.cpp CUDAFull GPUFits*~220–284 t/s Estimated
96 GB workstation GPU + 256 GB RAMFP16llama.cpp CUDAFull GPUFits*~387–608 t/s Estimated
Mac, 24 GBFP16llama.cpp MetalUnified memoryFits*~81–97 t/s Estimated
Mac, 48 GBFP16llama.cpp MetalUnified memoryFits*~81–97 t/s Estimated
Mac, 128 GBFP16llama.cpp MetalUnified memoryFits*~132–201 t/s Estimated
Mac, 512 GBFP16llama.cpp MetalUnified memoryFits*~147–226 t/s Estimated
DGX Spark 128 GBFP16llama.cpp CUDAUnified memoryFits*~87–96 t/s Estimated
Ryzen AI Max+ 395, 128 GBFP16llama.cpp VulkanUnified memoryFits*~58–85 t/s Estimated
CPU only, 64 GB DDR5-5600 dual channelQ4_K_Mllama.cpp CPUCPU onlyFits*~26–31 t/s Estimated extrapolated

Same plan engine as the planner, with its default reserves and RAM. How plans are chosen.

Files exact Hugging Face bytes

ArtifactSizeFilesRevisionSource
FP16
LiquidAI/LFM2.5-1.2B-Instruct-GGUF
2.18 GiB18ed288026eHF files
Q8_0
LiquidAI/LFM2.5-1.2B-Instruct-GGUF
1.16 GiB18ed288026eHF files
Q5_K_M
LiquidAI/LFM2.5-1.2B-Instruct-GGUF
0.79 GiB18ed288026eHF files
Q4_K_M
LiquidAI/LFM2.5-1.2B-Instruct-GGUF
0.68 GiB18ed288026eHF files
MLX 8-bit 8-bit
LiquidAI/LFM2.5-1.2B-Instruct-MLX-8bit
1.16 GiB1ce5150fd0dHF files
MLX 4-bit 4-bit
LiquidAI/LFM2.5-1.2B-Instruct-MLX-4bit
0.61 GiB17ccafdb04cHF files
Native safetensors BF16
LiquidAI/LFM2.5-1.2B-Instruct
2.18 GiB10f604ada3fHF files

Not found in the representative quant search (not proven unavailable): EXL3, AWQ, GPTQ. Complete selected file sets; projectors, MTP/draft and lookup files are listed separately and budgeted only by plans that use them. File bytes are not allocated memory.

Quality

AA evaluated modeIndexLiveCodeBenchTerminal-Bench 2.1GPQAHLESource
LFM2.5-1.2B-Instruct AA estimate5.20.0%—32.6%6.7%source

Arena text rating

Unknown

Publisher-reported benchmarks

None extracted from the model card.

AA and Arena describe the hosted model in the stated mode, not each quantized file; scores are never inherited by base, pruned, abliterated or quantized variants and never averaged together.

How to run it

llama.cpp · CPU/Metal/CUDA according to installed llama.cpp build

publisher/backend-documented, not executed source

uv pip install huggingface-hub
hf download LiquidAI/LFM2.5-1.2B-Instruct-GGUF lfm2.5-1.2b-instruct-q4_k_m.gguf --local-dir .
llama-server -m lfm2.5-1.2b-instruct-q4_k_m.gguf -c 4096 --port 8080
llama-cli -m lfm2.5-1.2b-instruct-q4_k_m.gguf -c 4096 --color -i \
    --temp 0.1 --top-k 50 --repeat-penalty 1.05

Prerequisites: See linked source for backend installation/version requirements.

Sampling: temperature=0.1 · top_k=50 · repeat_penalty=1.05

  • Exact local q4_k_m filename avoids default-hf artifact ambiguity.
  • 4096 context is documented example, not optimal universal value.

vLLM · unknown

publisher/backend-documented, not executed source

uv pip install -U vllm
vllm serve LiquidAI/LFM2.5-1.2B-Instruct

Prerequisites: See linked source for backend installation/version requirements.

  • Latest vLLM is unpinned; verify loaded architecture and requested context before using fit results.

Recipes are source examples (settings are the publisher’s or author’s, not measured optima) and were not executed by LocalFit. A downloadable file does not prove your backend build supports the architecture.

Reported speeds 0

No reported measurement yet. Submit one.