readygpu

LFM2.5 8B-A1B

8.5B total, 1B active · 125k context · lfm1.0 · released 2026-05-28

chat

What to run it on in Germany

Fully in memory at 8k context, priced at retailers today.

Something we price holds it fully in memory, but none reaches 15 tok/s, our bar for recommending it. The table below has the speeds.

All hardware that runs it

Everything we price, best value first. Two cards split the model by layers.

HardwareQuantFitsDecodePricetok/s per 1,000
NVIDIA GeForce RTX 3090Q8_0fits237 tok/s––
NVIDIA GeForce RTX 5090Q8_0fits455 tok/s––
Mac Studio M4 Max 128GBQ8_0fits145 tok/s––
NVIDIA GeForce RTX 4090Q8_0fits256 tok/s––
NVIDIA GeForce RTX 5070 TiQ8_0fits227 tok/s––
Mac mini M4 Pro 64GBQ8_0fits73 tok/s––
NVIDIA GeForce RTX 5060 Ti 16GBQ8_0fits114 tok/s––
AMD Radeon RX 7900 XTXQ8_0fits216 tok/s––
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits58 tok/s––
Mac Studio M3 Ultra 512GBQ8_0fits218 tok/s––
NVIDIA DGX Spark 128GBQ8_0fits69 tok/s––
NVIDIA GeForce RTX 3060 12GBQ8_0fits91 tok/s––
AMD Radeon RX 9070 XTQ8_0fits144 tok/s––
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits58 tok/s––
Mac Studio M3 Ultra 256GBQ8_0fits218 tok/s––
Mac mini M4 Pro 48GBQ8_0fits73 tok/s––
NVIDIA GeForce RTX 4060 Ti 16GBQ8_0fits73 tok/s––
NVIDIA GeForce RTX 5080Q8_0fits244 tok/s––
Mac Studio M4 Max 64GBQ8_0fits145 tok/s––
NVIDIA GeForce RTX 4070 Ti SuperQ8_0fits170 tok/s––
NVIDIA GeForce RTX 5070Q8_0fits170 tok/s––
AMD Radeon RX 9060 XT 16GBQ8_0fits72 tok/s––
Mac Studio M3 Ultra 96GBQ8_0fits218 tok/s––
NVIDIA GeForce RTX 4070 SuperQ8_0fits128 tok/s––
NVIDIA GeForce RTX 4080 SuperQ8_0fits187 tok/s––
NVIDIA RTX PRO 6000 BlackwellQ8_0fits455 tok/s––
AMD Radeon AI PRO R9700Q8_0fits144 tok/s––
AMD Radeon RX 7800 XTQ8_0fits140 tok/s––
AMD Radeon RX 7900 XTQ8_0fits180 tok/s––
ASUS Ascent GX10 128GBQ8_0fits69 tok/s––
Framework Desktop Ryzen AI Max+ 395 64GBQ8_0fits58 tok/s––
Intel Arc B580Q8_0fits86 tok/s––
Mac mini M4 32GBQ8_0fits32 tok/s––
NVIDIA GeForce RTX 3090 TiQ8_0fits256 tok/s––
AMD Radeon RX 7600 XTQ8_0fits65 tok/s––
Mac mini M4 16GBQ8_0fits32 tok/s––
NVIDIA GeForce RTX 3080 10GBUD-Q6_Kfits228 tok/s––
Intel Arc Pro B60Q8_0fits86 tok/s––
NVIDIA Tesla P40Q8_0fits88 tok/s––
Intel Arc A770 16GBQ8_0fits105 tok/s––

Memory needed

Weights, KV cache and runtime buffers at each context length (KV cache at fp16).

QuantFile@8k@32k@64k
UD-Q4_K_Mgguf5.0 GB6.1 GB7.3 GB9.0 GB
UD-Q5_K_Mgguf5.9 GB7.0 GB8.3 GB10.0 GB
UD-Q6_Kgguf6.6 GB7.7 GB9.0 GB10.7 GB
Q8_0gguf8.4 GB9.5 GB10.8 GB12.5 GB

Run it

Copy a command for the runtime you use.

  • UD-Q4_K_M
    • Ollamaollama run hf.co/unsloth/LFM2.5-8B-A1B-GGUF:UD-Q4_K_M
    • LM Studiolms get unsloth/LFM2.5-8B-A1B-GGUF@ud-q4_k_m
    • llama.cppllama-server -hf unsloth/LFM2.5-8B-A1B-GGUF:UD-Q4_K_M
  • UD-Q5_K_M
    • Ollamaollama run hf.co/unsloth/LFM2.5-8B-A1B-GGUF:UD-Q5_K_M
    • LM Studiolms get unsloth/LFM2.5-8B-A1B-GGUF@ud-q5_k_m
    • llama.cppllama-server -hf unsloth/LFM2.5-8B-A1B-GGUF:UD-Q5_K_M
  • UD-Q6_K
    • Ollamaollama run hf.co/unsloth/LFM2.5-8B-A1B-GGUF:UD-Q6_K
    • LM Studiolms get unsloth/LFM2.5-8B-A1B-GGUF@ud-q6_k
    • llama.cppllama-server -hf unsloth/LFM2.5-8B-A1B-GGUF:UD-Q6_K
  • Q8_0
    • Ollamaollama run hf.co/unsloth/LFM2.5-8B-A1B-GGUF:Q8_0
    • LM Studiolms get unsloth/LFM2.5-8B-A1B-GGUF@q8_0
    • llama.cppllama-server -hf unsloth/LFM2.5-8B-A1B-GGUF:Q8_0