I stopped using Qwen and Gemma after finding a local LLM that actually thinks before answering

Published Jul 24, 2026, 1:00 PM EDT Nolen began their writing career in 2019, with three years dedicated to editing the Creative section at MakeUseOf. Their expertise lies at the crossroads of technology and creativity, covering areas like photography, video editing, and graphic design. Outside of work, you'll often find Nolen diving into a good book, writing their own stories, or playing video games. Most local setups run fine on two or three solid generalists splitting the work between them. Qwen and Gemma both handle most of my day-to-day tasks, so I always tend to default to them. However, confining yourself to just one or two models is not the best approach to local LLMs, especially if you're on regular hardware that can't support the huge models that can handle everything you throw at it. It's better to split the work by task among several models. There was a model I kept putting off trying, mostly because I didn't fully get what it actually was. DeepSeek R1 Distill kept showing up in every local LLM conversation I was following, and at first it was hard to see what made it so different from models I was already running. But it turned out to be the specialist I didn't know I was missing… Want to stay in the loop with the latest in AI? The XDA AI Insider newsletter drops weekly with deep dives, tool recommendations, and hands-on coverage you won't find anywhere else on the site. Subscribe by modifying your newsletter preferences! What DeepSeek R1 Distill actually is And it confused me at first DeepSeek built its reputation on R1, a reasoning model trained mostly through reinforcement learning instead of the supervised-heavy approach most labs still lean on, and it landed near OpenAI's o1 on math and coding benchmarks, which is what got people outside the usual research circles paying attention. What confused me is that R1 itself isn't really the model most people are running locally. DeepSeek took the reasoning R1 produces, generated around 800,000 examples of it thinking through problems, and used that to retrain a handful of existing open models instead of shipping one single distilled product. That's why you'll see a Llama version and a Qwen version floating around with the exact same DeepSeek name attached, which is what tripped me up the first time. Basically, it's like having the same teacher but different students. Six sizes came out of that process, from 1.5B up to 70B, split across Llama and Qwen bases, and which one you grab actually changes what you're getting, not just how fast it runs. Llama 8B has the edge on code-heavy benchmarks, though Qwen 7B pulls ahead once math and general reasoning come into it, so grabbing whichever one happens to say "DeepSeek" doesn't tell you much by itself. Putting the model to the test Where it proved itself useful I ended up going with DeepSeek R1 Distill Llama 8B, and my first test ride was a riddle: the classic farmer-has-17-sheep one where the trick sits entirely in the wording. The thinking block it produced ran long, re-deriving its own logic with a smaller test case just to check itself before settling on an answer, and it got there correctly in the end, though I'm still not sure if that's thorough or just excessive for such a simple riddle. A riddle's low stakes though, so the real test was messier stuff and work that doesn't have one clean answer. I pulled raw markup straight out of a live site's dev tools and asked for a Nielsen's heuristics evaluation, and what stood out wasn't the verdict so much as watching it check actual class names and styles against each heuristic instead of reaching for whatever generic answer usually shows up first. It flagged the Sign Up button as a visibility issue, called out the paragraph-tag styling making it read as body text instead of an action, and worked through navigation and consistency without me steering it. The stronger example came from three user UX complaints I fed it, which didn't connect on their own, and the ask was one root cause instead of three separate fixes. It worked through each complaint individually first, then circled back and tested a few competing explanations before landing on one, and that circling-back part is basically the whole case for why the visible thinking matters at all. You're not just getting an answer, you're watching it rule out weaker ones first. You need more than one model DeepSeek R1 Distill is just another specialist to add to the stack None of this means Qwen and Gemma are getting uninstalled, and that's not really the argument I'm making anyway since both are still doing most of the daily work. The actual issue is hardware. Nobody running a single consumer GPU is finding one 70B model that covers everything well, mostly because that tier barely exists at sizes people can actually run at home. So the smarter move is matching the model to whatever job's sitting in front of you, rather than hunting for one that does everything. Gemma covers its usual ground, Qwen picks up plenty on its own, and this distill slots into whatever's left, mainly the stuff where watching the reasoning matters as much as the answer. My own setup already backs this up, since I'm not running one model for everything, I'm running several, and each one earns its spot for a different reason. Which one to reach for, and when, turned out to be the actual skill here. My local stack keeps growing I think this one earns a permanent spot, though not by pushing anything else out. Qwen and Gemma are still doing the bulk of what I need from a local model. What DeepSeek R1 Distill gives me is a model I actually trust for the reasoning-heavy tasks, which is a lane the other two can get a bit shaky in.