Liquid AI releases LFM2.5-8B-A1B

Posted by PauLabartaBajo@reddit | LocalLLaMA | View on Reddit | 32 comments

Liquid AI released LFM2.5-8B-A1B, an edge model designed to power real-life applications.

It builds on LFM2-8B-A1B with three major upgrades: an expanded 128K context window, 38T tokens of pre-training (up from 12T), and large-scale reinforcement learning. It also comes with a doubled vocabulary to improve tokenization for non-Latin languages.

The result is a model that chains tool calls, completes complex tasks, and fits comfortably on an entry-level laptop.

The model is available on HF > https://huggingface.co/LiquidAI/LFM2.5-8B-A1B

[-]

drooolingidiot@reddit

Looks good, but kind of strange that they're comparing against Qwen3 when Qwen3.6 of the same size exists.

[-]

StripperCunt@reddit

Qwen3.5 8B dense also exists

[-]

SirRece@reddit

Same language errors. God i hate the current internet.

[-]

MrPecunius@reddit

You must have auto-translate turned on: the second post was originally in Portuguese.

[-]

SirRece@reddit

Oh interesting, I didnt even know that was a setting. Awesome, TIL

[-]

Edenar@reddit

it's a 17GB model with around 2GB active param, so it's similar in size to gpt-oss-20b. But if it's better for tool call i'll give it a try.

[-]

Middle_Bullfrog_6173@reddit

It's similar size to gpt-oss 20b if you compare bf16 vs mxfp4. It's much smaller even in Q8 (9GB).

[-]

Edenar@reddit

i compare them because gpt-oss-20b is benched at mxfp4 (it's the official release) and lfm is benched at BF16. So it can give a false impression that a 8B model is competing with another 20B model when the 20B benched is actually smaller in size !

[-]

PhilippeEiffel@reddit

If you consider LFM2.5-8B-A1B in Q8, yes it is similar size to gpt-oss-20b. But then you have to run benchmarks in Q8 to compare results...

[-]

Middle_Bullfrog_6173@reddit

No, it's much smaller in Q8 than gpt-oss in mxfp4. But yes, third party benchmarks needed, even for bf16.

[-]

PhilippeEiffel@reddit

gpt-oss-20b native size is 10 GB. So LFM2.5-8B-A1B is 70 % bigger at full size (the size used for benchmarks).

[-]

Tall-Ad-7742@reddit

I may be wrong but i don't remember GPT-OSS-20B being just 10GB?! What version exactly are you talking about?

[-]

PhilippeEiffel@reddit

gpt-oss-20b is a 21B model A3.6B

It is native MXFP4 and it's size is 12 GB (https://huggingface.co/ggml-org/gpt-oss-20b-GGUF/tree/main)

So, it is not 10 GB, you are right!

[-]

Tall-Ad-7742@reddit

well it wasn't even meant that way i thought i took closer to 20GB but i guess i was atleast somewhat wrong but thanks for the explanation 👍

[-]

chille9@reddit

Now compare to qwen 3.6.

[-]

Sufficient-Bid3874@reddit

Holy benchmax

[-]

Every-Walrus@reddit

no not really, all these benchmarks are either tool calling, hallucination rate or instruction following. all things that are bound to how much post training the model got. LFM2.5 seems to be a post trained version of lfm 2 so it doesn't surprise me that it is good at these benchmarks. you can see a similar thing with qwen3.5 to qwen3.6

[-]

Glittering_Focus1538@reddit

There's already apex versions of this available, you can run this on 3 gb's of vram T_T