--- title: Liquid AI's LFM2.5-2.6B runs local agents on phones and... description: Liquid AI has released LFM2.5-2.6B, a 2.6B parameter model designed for on-device agentic tasks. It supports tool calling and multi-step workflows while... canonical: https://daily.dev/posts/liquid-ai-s-lfm2-5-2-6b-runs-local-agents-on-phones-and-laptops-vspjopwi8 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Liquid AI's LFM2.5-2.6B runs local agents on phones and laptops | daily.dev og:description: Liquid AI has released LFM2.5-2.6B, a 2.6B parameter model designed for on-device agentic tasks. It supports tool calling and multi-step workflows while... og:url: https://daily.dev/posts/liquid-ai-s-lfm2-5-2-6b-runs-local-agents-on-phones-and-laptops-vspjopwi8 og:image: https://api.daily.dev/og/posts/VsPjopwi8.png og:image:alt: Liquid AI's LFM2.5-2.6B runs local agents on phones and laptops og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Liquid AI's LFM2.5-2.6B runs local agents on phones and laptops **[Collections](https://daily.dev/sources/collections)** · 2 min read · 1 upvotes · 0 comments ## Summary Liquid AI has released LFM2.5-2.6B, a 2.6B parameter model designed for on-device agentic tasks. It supports tool calling and multi-step workflows while staying under 2.5 GB of memory. Performance benchmarks show 220 tokens/s on Apple M5 Max, 113 tokens/s on AMD Ryzen, and ~30 tokens/s on phones. Training used supervised fine-tuning, multi-domain distillation from specialist teacher models, and agentic reinforcement learning in real agent harnesses. On instruction following and tool use, it competes with models up to 4x its size, though larger models still lead on coding tasks. The model ships with day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX, and is available on Hugging Face. ## Content Liquid AI released [LFM2.5-2.6B](https://huggingface.co/LiquidAI/LFM2.5-2.6B), a 2.69B parameter hybrid model aimed at running agentic workloads directly on consumer hardware. The architecture is a bit unusual: 30 layers total, mixing 22 short convolution blocks with 8 grouped-query attention layers. It was pre-trained on roughly 34 trillion tokens and fine-tuned through a combination of supervised learning, multi-domain distillation from larger specialist models, and reinforcement learning inside actual agent harnesses. ## What it can do The model supports tool calling, multi-step reasoning, RAG, and long-context tasks up to 128K tokens. On the speed side, it hits 220 tokens/s on an Apple M5 Max, 113 tokens/s on an AMD Ryzen AI Max+ 395, and around 30 tokens/s on phones — all while fitting in under 2.5 GB of memory. On a single H100 at high concurrency, it reaches roughly 15K output tokens per second. Liquid claims it competes with models up to 4x its size on instruction following and tool use benchmarks. Day-one inference support covers llama.cpp, MLX, vLLM, SGLang, ONNX, and LM Studio. Fine-tuning is available through Unsloth and TRL. ## Where it falls short Liquid is upfront that agentic coding and knowledge-heavy tasks aren't where this model shines — larger models hold a clear advantage there. Early hands-on testing from at least one developer found it underperforming compared to other sub-3B dense models like Nanbeige4.2-3B across multiple tasks, suggesting the benchmark numbers may not fully translate to real-world use cases yet. The speed is genuinely impressive for the size. Whether it's ready for production agentic workloads probably depends on what those workloads actually involve — tool calling and data extraction look like reasonable fits, while anything requiring deep coding or broad factual knowledge probably still needs a bigger model. ## Questions this post answers ### How fast does Liquid AI LFM2.5-2.6B run on Apple M-series chips? LFM2.5-2.6B hits 220 tok/s on an Apple M5 Max and 200+ tok/s on an M4 Max. Memory usage stays under 2.5 GB, making it fit comfortably on most modern consumer devices. On AMD Ryzen it reaches 113 tok/s, and on phones it runs at around 30 tok/s. _Developers evaluating on-device LLM speed across hardware follow new model releases like this on daily.dev._ ### What inference frameworks does Liquid AI LFM2.5-2.6B support at launch? LFM2.5-2.6B launched with day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX. The model is available on Hugging Face, and a fully in-browser demo has already been built using it, illustrating how lightweight it is. _Teams choosing an inference stack for small on-device models track compatibility news like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 58 comments across hackernews (as of 2026-08-11). **TL;DR:** The community is cautiously interested in LFM2.5-2.6B's edge/on-device potential and non-coding agentic use cases, but tempers enthusiasm with skepticism about benchmark cherry-picking, real-world coherence issues, and missing comparisons to strong competitors like Qwen. **Sentiment:** 40% positive · 35% mixed · 25% skeptical **The case for** - The hybrid architecture targets reliable tiny-model operation in ways other model families don't, with models as small as 230M params. - Practical non-coding agentic workflows (scheduling, ops runbooks, home assistants, data extraction) are seen as genuinely useful fits for this class of model. - Runs well on existing consumer hardware without dedicated GPU, making local deployment accessible. **The pushback** - Multiple commenters report poor real-world coherence and tool-calling behavior compared to other models. - Benchmark comparisons are suspected of cherry-picking, with notable omissions like Qwen3.5 2B and Nanbeige4.2-3B. - The claim of being competitive with models 4x its size is questioned, with at least one commenter saying it doesn't beat a 2x-sized Qwen 4B. - Benchmarks lack clear 'higher/lower is better' labels, making them hard to interpret. **By community** - hackernews (mixed): Genuine excitement about edge agentic use cases is offset by skepticism over cherry-picked benchmarks, real-world coherence problems, and missing competitor comparisons. **Hottest debate:** Whether LFM2.5-2.6B's benchmark claims hold up against omitted competitors like Qwen, or whether the self-reported comparisons are cherry-picked. **Open questions** - How does it actually compare to Qwen3.5 2B and other similarly-sized models not included in the benchmark table? - What specific agentic (non-coding) workflows and tooling environments work best with this model in practice? **Highlights** > LFM's training/post-training is famously different than other models. They target reliable operation of tiny models in ways other model families don't (they aren't just scaling a larger model to a smaller size). If you're looking for good performance out of tiny models, LFM has the most advanced design. Note how they're much smaller than all other models in the comparison yet match or exceed them. This is for 2.6B params, but they have models as small as 230M. Nobody else designs models that small. > — [0xbadcafebee on hackernews · 2 comments](https://news.ycombinator.com/item?id=49253806) > > Note how they're much smaller than all other models in the comparison yet match or exceed them. There's a strong incentive to cherry pick in self-reported comparisons. If there is a model that's better, it gets left out. Have you seen Nanbeige4.2-3B or Ling-3.0-tiny? > Nobody else designs models that small. There are people building even smaller models. > — [woadwarrior01 on hackernews](https://news.ycombinator.com/item?id=49254201) > It's not even competitive with 2x sized Qwen 4B. Why is Qwen3.5 2B not in the table? > — [lostmsu on hackernews](https://news.ycombinator.com/item?id=49254399) > Yeah. Clearly the target for these models is on-device assistant/analyst tasks where really everything is agentic. Thinking out loud: any text-based generative AI application is on a continuum between: - document-to-document: solve this language problem using language skills by outputting only new symbols, and - document-to-tasks: solve this language problem by only operating these tools. In a sense, given access to vast compute to train a very large model, agentic coding is one of the easier applications that is somewhere between the two. It is manipulating symbols in a mixture of languages that are biased towards context-free (code), it exploits the embedded knowledge in a vast number of weights, and it calls fairly simple tools. The user-focussed solution is happening inside the LLM. A small model can't do this job well, even if it has a good understanding of programming languages, because it lacks the world knowledge to understand the problem. At the other end of the continuum is: these words mean do this thing. That Cactus Needle 2 model mentioned earlier is here. It doesn't even produce a language description of what it did, because it one-shots tasks. In the middle but near this end is the fantastically hard job: train a model that understands language and reasons well enough to respond to queries about the tools it has access to, operate them and reply in natural language, without being large or slow. That is, understand language without being overburdened by details of scenarios that caused the need for the words in the first place. This is maybe the great grand challenge of LLMs: make it know how to speak and reason and think and work for us while discarding everything that is just knowledge of unnecessary facts. Small reasoning models are where most of the really big challenges are. > — [dofm on hackernews](https://news.ycombinator.com/item?id=49254602) > These LiquidAI models have never worked well for me in practice. > — [Gecko4072 on hackernews · 2 comments](https://news.ycombinator.com/item?id=49253590) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=49173107) · 100 points · 58 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#local-ai](https://daily.dev/tags/local-ai), [#llama-cpp](https://daily.dev/tags/llama-cpp) [View this post on daily.dev](https://daily.dev/posts/liquid-ai-s-lfm2-5-2-6b-runs-local-agents-on-phones-and-laptops-vspjopwi8) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Liquid AI's LFM2.5-2.6B runs local agents on phones and laptops","url":"https://daily.dev/posts/liquid-ai-s-lfm2-5-2-6b-runs-local-agents-on-phones-and-laptops-vspjopwi8","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/liquid-ai-s-lfm2-5-2-6b-runs-local-agents-on-phones-and-laptops-vspjopwi8"},"datePublished":"2026-08-04T18:44:01.298Z","dateModified":"2026-08-11T13:33:48.963Z","description":"Liquid AI has released LFM2.5-2.6B, a 2.6B parameter model designed for on-device agentic tasks. It supports tool calling and multi-step workflows while...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/liquid-ai-s-lfm2-5-2-6b-runs-local-agents-on-phones-and-laptops-vspjopwi8","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents,local-ai,llama-cpp","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Liquid AI's LFM2.5-2.6B runs local agents on phones and laptops"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/liquid-ai-s-lfm2-5-2-6b-runs-local-agents-on-phones-and-laptops-vspjopwi8#faq","mainEntity":[{"@type":"Question","name":"How fast does Liquid AI LFM2.5-2.6B run on Apple M-series chips?","acceptedAnswer":{"@type":"Answer","text":"LFM2.5-2.6B hits 220 tok/s on an Apple M5 Max and 200+ tok/s on an M4 Max. Memory usage stays under 2.5 GB, making it fit comfortably on most modern consumer devices. On AMD Ryzen it reaches 113 tok/s, and on phones it runs at around 30 tok/s. Developers evaluating on-device LLM speed across hardware follow new model releases like this on daily.dev."}},{"@type":"Question","name":"What inference frameworks does Liquid AI LFM2.5-2.6B support at launch?","acceptedAnswer":{"@type":"Answer","text":"LFM2.5-2.6B launched with day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX. The model is available on Hugging Face, and a fully in-browser demo has already been built using it, illustrating how lightweight it is. Teams choosing an inference stack for small on-device models track compatibility news like this on daily.dev."}}]} ```