How I Tested 5 Small LLMs on a Weak PC (Intel i5, No GPU) – And Found a Winner

A developer tested five small LLMs (under 2B parameters) on a budget PC with an Intel i5 CPU, no GPU, and single-channel RAM. The LFM2.5-1.2B-Instruct model emerged as the best all-around performer for CPU-only systems, offering the best balance of speed and quality. The test highlighted that single-channel RAM creates a memory bandwidth bottleneck limiting performance.

A practical guide to running LLMs on budget hardware: real speeds, real stories, and real conclusions Before anything, let me show you what I was working with. No GPU. No high-end hardware. Just a regular office PC: | Component | Specification | |---|---| CPU | Intel Core i5-10400 @ 2.90GHz 6 cores | RAM | 16GB DDR4 Single Channel – important | GPU | Intel UHD Graphics 630 128MB – basically useless | Storage | 238GB SSD | Software | LM Studio GGUF format models | Key limitation: The single-channel RAM creates a memory bandwidth bottleneck of ~20 GB/s. This is the real reason speeds can't go much higher. Most LLM benchmarks and reviews assume you have: But what if you have none of that ? What if you're a developer on a budget, a student, or someone with an old PC? I wanted to find the best small LLM under 2B parameters that actually runs well on hardware like mine. No theory. Real tests. Real speeds. Real stories. And yes, I made each model write a funny cat story to test creativity, coherence, and humor. | | Model | Size | Format | |---|---|---|---| | 1 | LFM2.5-350M | 350M | GGUF Q4 K M | | 2 | Qwen3 0.6B Instruct | 0.6B | GGUF Q4 K M | | 3 | LFM2.5-1.2B-Instruct | 1.2B | GGUF Q4 K M | | 4 | Gemma-3-1B-Uncensored | 1B | GGUF Q4 K M | | 5 | DeepSeek-R1-Distill-Qwen-1.5B | 1.5B | GGUF Q4 K M | For each model, I did: This was the fastest model by far . The response appeared almost instantly. "Milo the cat lived in Milo's tiny, fussy home. One sunny afternoon, he'd tried to sneak into the kitchen for coffee—only to be caught by a curious squirrel named Sammy..." | Aspect | Score | Notes | |---|---|---| | Coherence | 7/10 | Mostly logical, but names got confusing "Milo" = cat AND owner | | Humor | 6/10 | Tried hard but felt forced | | Originality | 7/10 | Creative premise a cat wanting coffee | Verdict: Perfect for summarization and quick tasks. Not ideal for creative writing. Solid speed. A noticeable step down from 350M, but still very responsive. "Whiskers wasn't your average cat—he had a knack for solving puzzles faster than you could say 'purr'..." | Aspect | Score | Notes | |---|---|---| | Coherence | 7/10 | Decent structure, no major confusion | | Humor | 6/10 | Acceptable but predictable | | Originality | 6/10 | Standard "clever cat" tropes | Verdict: A solid general-purpose model. Nothing special, but nothing broken. The slowest of the "good" models, but the quality jump was worth it. "Once upon a time, in a quirky little town named Pawsville, lived a fluffy gray tabby cat named Whiskers. Whiskers wasn't your average cat—he had a knack for solving puzzles faster than you could say 'purr'... In this magical realm, animals were talking animals—dogs with tiny glasses, birds with tiny hats, even a wise old owl who wore a monocle..." | Aspect | Score | Notes | |---|---|---| | Coherence | 9/10 | Excellent structure from beginning to end | | Humor | 8/10 | Genuinely funny "owl with a monocle," "squirrel trying to juggle carrots" | | Originality | 8/10 | Rich world-building, consistent characters | Example of good humor: "He met a grumpy old turtle named Timmy, who kept guarding a treasure chest filled with shiny seashells. The turtle was so stubborn, he'd stare at Whiskers for hours, refusing to let him in." Verdict: The best all-around model for CPU-only systems. Use this for chat, story writing, summarization, and daily tasks. The slowest, but with a unique personality. Interesting behavior: The model "thought" for 1 minute 27 seconds before responding. This is likely due to its uncensored nature exploring multiple response candidates. "Mittens squeezed her eyes shut and jumped right into the hole. She tumbled down a dark slope, landing in a pile of old magazines and a bag of catnip she had been hiding behind the sofa for later... The human just laughed, shook their head and said: 'That's why I left my laptop open.'" | Aspect | Score | Notes | |---|---|---| | Coherence | 7/10 | Slightly chaotic but entertaining | | Humor | 8/10 | Dry, adult-oriented humor. The punchline was genuinely unexpected | | Originality | 8/10 | Very unique voice | Verdict: Great for personal entertainment if you want a different flavor of humor. Too slow for daily use. Thought for 33 seconds before responding. This is a "reasoning model" designed for math and logic, not storytelling. "Uh-oh exclaimed a neighboring neighbor, Squidward... Uh-oh, he said again... Uh-oh, Whiskers said again... Uh-oh, Whiskers said once more..." | Aspect | Score | Notes | |---|---|---| | Coherence | 3/10 | Extremely repetitive, characters appear/disappear randomly | | Humor | 2/10 | "Uh-oh" repeated ~15 times is not funny | | Originality | 4/10 | Some creative elements but lost in chaos | Verdict: Do not use for creative writing. This model is for math, logic, and step-by-step reasoning. I misused it, and the results show why. | Rank | Model | Speed t/s | Coherence | Humor | Best For | |---|---|---|---|---|---| | 🥇 | LFM2.5-1.2B-Instruct | 13.5 | 9/10 | 8/10 | Everything chat, stories, summarization | | 🥈 | LFM2.5-350M | 36 | 7/10 | 6/10 | Fast summarization, always-on assistant | | 🥉 | Qwen3 0.6B | 20 | 7/10 | 6/10 | General-purpose backup | | 4 | Gemma-3-1B-Uncensored | 10 | 7/10 | 8/10 | Personal entertainment adult humor | | 5 | DeepSeek-R1-Distill-Qwen-1.5B | 10.4 | 3/10 | 2/10 | Math/logic NOT stories | The 350M model was 3x faster than the 1.2B Instruct, but the story quality was noticeably lower. LFM2.5-350M 350M params outperformed Qwen3 0.6B 600M params in multiple benchmarks. DeepSeek-R1 is amazing at math but produces repetitive, incoherent stories. Use the right tool for the right job. Models larger than 1.5B drop below 10 t/s on my hardware. Models smaller than 1B sacrifice too much quality. They consistently outperformed competitors in both speed and quality on my Intel i5. 👉 LFM2.5-1.2B-Instruct 👈 You don't need a $2000 GPU to run LLMs locally. With a humble Intel i5, 16GB RAM, and no graphics card, you can run LFM2.5-1.2B-Instruct at ~13 tokens/second and get genuinely useful results for: The models are getting smaller, faster, and smarter. LFM2.5 proves that 1.2B parameters can deliver quality that rivals larger models. Go try it yourself. Download LM Studio, grab the LFM2.5-1.2B-Instruct GGUF file, and start experimenting. The tests I ran were focused on a single, simple scenario: generating a funny cat story on a specific hardware setup. While this gave me clear, comparable results across five models, it's important to remember that LLM performance can vary significantly depending on the task. A model that writes a decent story might struggle with code generation, mathematical reasoning, or multi-turn conversations. Likewise, your hardware, software version, quantization settings, and even the phase of the moon okay, maybe not that last one can affect speeds and output quality. So take my findings as a useful data point, not a universal truth. You can also test models on your own workloads before making a decision.