LFM2.5-24B-A2B by Liquid AI
Ranked 23 of 39 open-weight language models by recommenders
Fast, efficient open-weight model for on-device agents; critics say it lags Qwen on knowledge and coding, and the license is confusing.
Fast, efficient open-weight model for on-device agents; critics say it lags Qwen on knowledge and coding, and the license is confusing.
33 people gave an opinion
- recommend it
- 23
- criticize it
- 10
- approval share
- 70%
Recommended in 2 of 13 months
- November 2025: not recommended
- December 2025: not recommended
- January 2026: not recommended
- February 2026: not recommended
- March 2026: not recommended
- April 2026: not recommended
- May 2026: recommended
- June 2026: not recommended
- July 2026: not recommended
- August 2026: not recommended
- September 2026: not recommended
- October 2026: recommended
- November 2026: not recommended (month in progress)
Why it doesn’t carry the badge
- 23 people recommend itNeeds 10 or more, met.
- 2.3 recommenders for every criticNeeds 2 to 1 or better, met.
- Recommended in 2 of 13 monthsNeeds 4 or more, not met.
The same three rules apply to every product, and the badge can’t be bought. How the badge works
What people say about it
The reasons people give, with the number of comments that give each one. Open a reason to read what they said.
Why people recommend it
Each mark is one comment
Fast output on modest hardware
“fast is fast (18tk/s on a Intel Core 7 155H laptop), but quality is something else”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” “I am getting at most 25tps with the llama cpp vulkan backend”
Comment from 2 May 2026 in “LFM2-24B-A2B: Scaling Up the LFM2 Architecture” “this thing puts out results very fast even on max reasoning”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” - Comment from 2 May 2026 in “LFM2-24B-A2B: Scaling Up the LFM2 Architecture”
- Comment from 2 May 2026 in “LFM2-24B-A2B: Scaling Up the LFM2 Architecture”
Good for on-device work
“Really Cool, the on-device shifts the economics a bit.”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” “It is a good model for on-device work.”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” “I just tried the excellent local model”
Comment from 15 Oct 2026 in “OpenCode 2.0” Strong results for its size
“Way more impressive than I expected. I am still messing about with it, but I think it's a keeper for general discussions.”
Comment from 26 Oct 2026 in “IBM CEO says open weight models are now good and cheap on-prem” “Wow, those are seriously impressive numbers for how intelligent this model seems to be!”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” “someone has caught up to Qwen (albeit Qwen from 1-2 months ago) in parameter efficiency”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” Handles terminal and scripting tasks
“it's stronger for terminal use, like chaining bash calls, monitoring background tasks, etc.”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” “these small agentic-focused models are a lot of fun to play with”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” “its doing pretty well writing some python scripts for me”
Comment from 26 Oct 2026 in “IBM CEO says open weight models are now good and cheap on-prem”
What people warn about
Each mark is one comment
Weak at coding and agentic loops
“Even Liquid's own documentation states that coding isn't this model's strong suit.”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” “kept going in circles fixing one compile error and introducing another”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” “So this wouldn't work in its current state in an agentic loop.”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” Behind Qwen on benchmarks
“it performs much, much worse than Qwen 35B A3B for general knowledge”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” “it appears to get a worse score on SWE-Bench-Verified (by a lot) when compared to Qwen 3.6 35B-A3B”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” Unclear non-commercial license clause
“this model has a questionably-enforceable non-commercial use clause in its license”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” “The license doesn't really make much sense to me”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” Fast but lower quality output
“fast is fast (18tk/s on a Intel Core 7 155H laptop), but quality is something else”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” “this is fast but maybe not usable”
Comment from 16 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents” “It is very fast on a 4090, but also very stupid.”
Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents”
Other open-weight language models people recommend
The four most recommended other open-weight language models, with the same counts for each, so you can compare like for like. The tick on each bar marks the 2 to 1 line the badge needs.
| Product | Recommend | Criticize | Approval | Months | Owned | Badge |
|---|---|---|---|---|---|---|
| Qwen 3.8 27B | 1,173 | 426 | 73% | 13 | 2 mo32 owners | Approved |
| Z.ai GLM | 1,043 | 384 | 73% | 13 | 2 mo52 owners | Approved |
| Kimi | 931 | 359 | 72% | 13 | 1 mo26 owners | Approved |
| Gemma 4 | 713 | 249 | 74% | 13 | 1 mo14 owners | Approved |
| LFM2.5-24B-A2B by Liquid AIThis page | 23 | 10 | 70% | 2 | — | — |
What people said
34 of the 36 comments that mention it, newest first. Each links to the original comment.
“Loved playing with LFM2.5”
“Way more impressive than I expected. I am still messing about with it, but I think it's a keeper for general discussions.”
“its doing pretty well writing some python scripts for me”
“this is fast but maybe not usable”
“Really Cool, the on-device shifts the economics a bit.”
“It got stuck on a lot of tool calls.”
“I just tried the excellent local model”
“I told it to create a 3D model in Three.js and it failed.”
“speed increases like this are rarely free lunches”
“It is very fast on a 4090, but also very stupid.”
“The resulting pelican was underwhelming.”
“So this wouldn't work in its current state in an agentic loop.”
“The license doesn't really make much sense to me”
“fast is fast (18tk/s on a Intel Core 7 155H laptop), but quality is something else”
“It is a good model for on-device work.”
“The license is generrous (free under <$10m)”
“Even Liquid's own documentation states that coding isn't this model's strong suit.”
“this model has a questionably-enforceable non-commercial use clause in its license”
“it performs much, much worse than Qwen 35B A3B for general knowledge”
“it's stronger for terminal use, like chaining bash calls, monitoring background tasks, etc.”
“kept going in circles fixing one compile error and introducing another”
“they always undertrain the model”
“it appears to get a worse score on SWE-Bench-Verified (by a lot) when compared to Qwen 3.6 35B-A3B”
“this thing puts out results very fast even on max reasoning”
“someone has caught up to Qwen (albeit Qwen from 1-2 months ago) in parameter efficiency”
“I'd rather be bullish on a new American AI lab making it to first place among western open-weight models”
“these small agentic-focused models are a lot of fun to play with”
“Wow, those are seriously impressive numbers for how intelligent this model seems to be!”
“the previous LFM2 model (lfm-2-24b-a2b) didn't do well on my tests, it got only 1/20 questions/tasks right”
“I am getting at most 25tps with the llama cpp vulkan backend”
“one of the first models that run at ~20tps on my laptop”
“multi turn iteration was not working well”
“pretty cool if you don't have a GPU”
“achieved 0.3% accuracy and hallucinated fake crisis resources”
Counts come from 1,795,640 Hacker News comments, November 2025 to November 2026. How we count. Report a mistake.