I paid Claude roughly $20 a month for everyday questions, weather lookups, and the occasional weekend study guide for my kid. Then I read Nolen Jonker’s piece on XDA Developers, where they put three sub-2B (sub-2 billion parameter) local LLMs through the exact same kinds of tasks, and I had to ask a hard question: was I paying $20 a month for capability, or for the absence of a particular failure mode I had learned to be afraid of?
The short version: there is now a tiny local model that is good enough at agentic work (calling external tools on your behalf, like a weather API or a search engine) that it can replace a meaningful slice of what I was paying for. It runs on a Raspberry Pi. It is 1.2 billion parameters. It is called LFM2.5-1.2B-Instruct, and it is built by Liquid AI, a company you have probably not heard of. Here is what changed and what it cost me.
What I was actually using Claude for
To be honest about this, the bulk of my usage was three things: structured study guides for my kid’s weekend reading, real-time weather and fact lookups via a search connector, and short rewrites of work emails when I could not think of a polite way to say "no." None of these are coding tasks. None of them need 100K context windows (the maximum amount of text a model can read in one conversation, measured in tokens). None of them need the model to be smarter than a well-read high schooler.
Nolen’s framing lined up with mine. They wrote: "I keep coming back to Qwen 3.5 9B and Gemma 4 E4B, both running fine on my 8GB of VRAM" (the dedicated memory on a graphics card). They were already running mid-sized models. The question that pulled me in was the one they asked next: how far down can you go before a local model becomes useless? Their answer was the model I am now running. Mine matches.
The three under-2B models Nolen tested
Nolen picked three small open-weight models (models whose internal settings can be downloaded and run on your own hardware, unlike closed services like ChatGPT or Claude) that still claim to do real work:
- Gemma 4 E2B (Google, 5GB of RAM at 4-bit quantization, where 4-bit is a compression technique that shrinks the model by roughly 75% with minor quality loss). The "E" stands for "effective" parameters. It uses Per-Layer Embeddings to keep the active memory footprint low while the full weights live elsewhere. Designed to run on phones and laptops.
- Qwen 3.5 0.8B (Alibaba, under 1GB, Apache 2.0 licensed, multimodal meaning it handles both text and images). The smallest entry in the Qwen 3.5 line, released in March 2026. Uses a Gated DeltaNet architecture (a newer alternative to the standard transformer design that uses less memory for long conversations) that lets the 0.8B model stretch to a 262K context window, which is wild for a model this size.
- LFM2.5-1.2B-Instruct (Liquid AI, under 1GB of memory, 32K context, hybrid architecture combining multiplicative gates with grouped query attention, which is a different way of organizing how the model reads long text). Liquid is upfront in their docs that it is built for agentic tasks, data extraction, and RAG (Retrieval-Augmented Generation, a setup where the model looks up facts from a document or database before answering), and explicitly not for knowledge-intensive work or programming.
Nolen ran each model through two tests. First, a structured weekend study guide, where the prompt asked for a specific format with section headers, a self-check section at the end, no specific tool names, and no meta-conclusion. Second, a real-time weather query through a Brave Search connector, which is a small piece of software that lets the model call out to a search engine to fetch current information.
The results, with the spin the article title hid
Nolen’s headline was that only one model impressed them. The actual finding was more useful. On the structured task, Gemma 4 E2B was the most reliable. It kept the format intact, did not hallucinate course names or book titles, and hit all six required sections. Its catch was a planning-process leak, where the model prints its internal scratchpad into the response, which Nolen pinned on LM Studio, the desktop app they were using. Qwen 3.5 0.8B drifted badly on the structured task, sliding from "visual hierarchy" into "CSS grid, flexbox, sticky nav." It also violated the "do not name tools" rule. LFM2.5-1.2B failed the structured task outright, throwing the format out the window and returning four prose paragraphs.
On the real-time task, the rankings flipped. LFM2.5-1.2B called the Brave Search tool cleanly and came back with "about 18°C with partly cloudy skies," in one sentence, in the correct unit, within a couple of degrees of actual conditions. Gemma 4 E2B worked the first time but flunked a follow-up. Qwen 3.5 0.8B told Nolen Cape Town was 57°C, then doubled down when corrected.
The win is not what I expected. The model that lost the creative task is the one Nolen would actually trust with the kind of work it was built to do. Tiny models are not toys, they are specialists. LFM2-1.2B won the agentic test because Liquid built it for agentic work. Qwen 3.5 0.8B was a fine-tuning base (a small model designed to be customized for a specific job, not used as-is for general chat) and got punished for being used as a general-purpose chatbot. Gemma 4 E2B sat in the middle and was the best generalist.
The migration: what was hard and what was easy
Setting up LFM2.5-1.2B on my Mac took about 15 minutes. I downloaded the GGUF file (a common file format for running models on regular computers, including Macs) from Hugging Face, pointed LM Studio at it, and started asking questions. The model is small enough that the entire weights file sits in memory without me thinking about it, and inference (the act of the model actually generating an answer) is fast on the M-series chip I am running.
The hard part was the MCP connector. MCP stands for Model Context Protocol, which is a standard way for a model to call out to other tools, like a search engine or a weather API. Wiring up the Brave Search MCP through LM Studio required reading three different sets of docs, two of which were out of date. Once it was running, I sent the model the same weather query Nolen used, and it returned the right answer in the right unit on the first try. That was the moment I canceled the Claude subscription.
For the structured study guide task, I had to adjust my expectations. The model that wins on tool use loses on structure. For weekend study guides I now use Gemma 4 E2B, which sits at about 5GB of RAM and is still small enough to run on the same machine. The two models are different tools for different jobs. That is the actual shape of "free local AI in 2026," and it is less convenient than "one model does everything," but it is also free.
What I gave up
Three things, and they are real:
- I gave up the comfort of a model that is reliably right. Claude hallucinates less than a 1.2B parameter model. That is not a moral judgment, it is a fact about parameter count. When I ask Claude about a niche topic, I get a careful answer. When I ask LFM2.5-1.2B, I get a confident one that is sometimes wrong. I have to spot-check.
- I gave up long context. LFM2.5-1.2B has a 32K context window. Claude’s is 200K. If I want to paste in a 50-page document and ask questions about it, I cannot do that on the local model without chunking the document and giving up coherence. For my use case, which is short questions, this does not matter. If yours involves long documents, it will.
- I gave up the "thinking" step. Larger reasoning models (like Claude with extended thinking or OpenAI’s o-series) show you their work before answering. Tiny models either have it baked in or they do not. Gemma 4 E2B has a configurable thinking mode. Qwen 3.5 0.8B does not. LFM2.5-1.2B does not. If you want a model to "show its work," you have to ask for it in the prompt, and the answer is usually less satisfying than what a larger model produces by default.
Trade-offs
- Cost: $0 per month for the local models, versus $20/month for Claude Pro. $240/year saved if you only need the local-tier capability.
- Hardware: LFM2.5-1.2B runs on a Raspberry Pi 5 with 8GB of RAM. Gemma 4 E2B needs roughly 5GB of system RAM. Qwen 3.5 0.8B runs on phones.
- Setup time: first-time install is 15-30 minutes including the MCP wiring. If you skip the MCP and just run the chat interface, it is 5 minutes.
- Reliability: tiny models hallucinate more often. For factual questions, you still want a web search or a larger model. For tool-calling and short answers, the reliability is fine.
- Privacy: the data never leaves your machine. For parents, lawyers, doctors, or anyone who works with sensitive text, this is the actual win. Cloud AI services retain prompts by default, and most enterprise tiers are still under audit.
Bottom line
If you were paying for a chatbot mostly to look up the weather, draft a study guide, or rewrite the occasional work email, you can replace the $20/month Claude Pro with a $0/month local setup that includes LFM2.5-1.2B for tool-calling tasks and Gemma 4 E2B for structured writing. The setup is not turnkey. You will spend an afternoon getting LM Studio configured, you will wire up an MCP connector, and you will learn that tiny models are specialists, not generalists. The savings are real. The privacy win is real. The capability ceiling is lower than what you are used to, and you have to know where that ceiling is.