Why your local LLM feels dumber than it is

(forum.level1techs.com)

171 points | by felineflock 6 hours ago

12 comments

  • nullpoint420 1 minute ago
    At least I'd be in control of model quality vs. when Anthropic decides to randomly drop the quality of their offering
  • jonplackett 4 hours ago
    I just got qwen 3.8 27b mlx running on my Macbook Pro and honestly I’m pretty blown away by how not-dumb it is.
    • prettyblocks 4 hours ago
      My problem is how hot they run. I'm on an m4 pro. Do you have the same issue?
      • walrus01 6 minutes ago
        If you're going to try to use a mac laptop for long term inference you'll probably want to buy one of those $30 laptop fan cooling pads (sold for absurd sized 17 inch x86-64 gaming laptops) on Amazon. Otherwise you'll roast the battery over time and significantly hurt its cycle life and capacity. If you can get significant airflow under the laptop that'll help a lot.
      • akg_67 1 hour ago
        Macs fan control set to full blast and lifting the Mac above table for air circulation across bottom.

        I use M1 Max with qwen3.8 27b mlx. The gpu temp can reach easily to 95°C as fan doesn’t kick in automatically until 90°C. With Macs fan control at full blast and MBP off table, temp usually hover around 85°C.

        • Footprint0521 32 minutes ago
          Just curious, what is your ram and how many tokens per second do you get with that?
          • pram 17 minutes ago
            I am getting like ~25 tps with M5 Max 64GB. Unsloth Desktop and Qwen 3.8 27B Q6 KXL
      • jonplackett 4 hours ago
        It’s hot and also LOUD and runs the battery down quick.

        But I’m having a lot of luck just running things when I’m away from the computer and can leave it plugged in.

        It starts going weird (unreliable and slow) with context over 80k so you have to pick tasks one at a time and baby sit a lot more than Claude. But it really is very capable and feels like there’s an intelligence there to talk to. Maybe gpt-4 level clever?

        I have an m5 max 64gb and I think anything slower would be quite painful.

        • seanmcdirmid 3 hours ago
          I have an m3 max 64GB and it runs ok, I was playing around with it last night. Its reasoning is better, but the token/s still makes an MoE model (35B A3B) preferable so I switched back.

          Try using Goose instead of Claude's harness? Goose doesn't load as much context in at the start, so it might be more usable. I've definitely been able to get it past 128K, although I typically don't go beyond 70-80k. MoE might also be better at this.

      • lukan 4 hours ago
        I don't have the hardware but a often mentioned advice is to put your mac into energy saving mode - it still will work, a bit slower, but stays cool.
      • ls612 2 hours ago
        I tested Gemma 4 a couple months ago and yes it ramps the fans to max. Definitely a thing that should only be done on wall power.
      • downrightmike 4 hours ago
        Mineral oil bath?
        • datadrivenangel 1 hour ago
          just decent air cooling and you'll be okay. it will get up to 75/80C though for my M5 MBP.
    • alexchantavy 4 hours ago
      How many tok/s are you getting? What gen mbp?
      • mattdw 3 hours ago
        My M4 Pro 48GB gets about 13tok/s, in both 3.6 and 3.8 27b Qwens. Qwen A3B and Gemma get closer to 100tok/s from memory but the results are pretty poor for coding tasks.

        Edited to add: for agentic workflow I’m running omlx which tells me it has about a 90% cache hit rate (tradeoff is some disk and mem space) - that noticeably changes the felt speed.

      • MrScruff 3 hours ago
        I get around 20 tok/s, 4 bit quant, MTP, 4 bit KV cache quantisation. On an M4 Pro 48Gb.
      • a11r 37 minutes ago
        M3 Pro 36GB. I am getting 17 tps with MTPLX.
      • dominotw 3 hours ago
        i suspect ppl dropping generic "its awesome" comments are not actually using it and prbly just managed to get it running for a prompt or two.
        • FireCrack 1 hour ago
          I feel like it's 50/50 between people doing that, and people that have spent a lot of time tuning a system they are pointing at focused and well specified problems.
        • petcat 2 hours ago
          Yeah, that's my experience. It's a big "wow" factor to get a non-trivial LLM running on my Mac, but it's actually not that useful. Like trying to use Photoshop at 8 FPS.
    • StarlaAtNight 4 hours ago
      how quick does it respond? what are specs of your laptop?
      • Gareth321 4 hours ago
        I tried it on my M1 MacBook Pro. It's slow but surprisingly smart as a general purpose LLM. Maybe GPT-5.3 level. I gave it a bunch of tools and it can search the internet, make product recommendations, document, code, etc.
        • alexpotato 3 hours ago
          Had the same reaction so had Grok create a script to:

          - find a free GPU droplet on digital ocean

          - fire it up

          - pull in a snapshot of the model + extra files/packages etc

          - set up a ssh tunnel so that the localhost:8000 routes to the above

          Then I just configured OpenCode to use the above and was off to the races.

          Works out to be about ~$2/hr all said and done which isn't bad as I only pay when I'm using it (but could get expensive with 24/7 running)

          • embedding-shape 3 hours ago
            > Works out to be about ~$2/hr all said

            What GPU you end up with for that price? Vast.ai (https://cloud.vast.ai/?priceInstanceHourlyMax=2) has a bunch of setups available to reach 192GB VRAM under $2 :) Quick skim showed 4x48, 2x96 and 8x24, all for under 2 buckaroos or around there.

          • b112 3 hours ago
            Awesome, I used Claude to write a small python script to do the same with Linode's API. The only difference is I setup a persistent drive, and with Linode you can boot off of it. So my biggest start up lag is ~ 2 minutes to deploy + boot, then maybe 2 more to warm the model.

            I actually dislike LLMs. But I'm a realist, and on-demand compute like this is massive cost saving measure.

            (persistent drives are relatively cheap, compared to a box with several GPUs.. or even one. I find it worth the expense)

      • chorlton2080 4 hours ago
        Does it need to respond fast? For important applications, I'm sure we'd all be fine waiting 20 minutes for a high quality, usable answer. Or is it the need for interative refinements that make speed relevant?
        • jonplackett 4 hours ago
          It requires patience but it’s more like waiting 5 mins for it to do tasks. You need to be much more involved though and do things slower than Claude where you can trust it to do a lot of tasks at once. It doesn’t have the context for that
        • dominotw 3 hours ago
          if you are so sure about what the final shape of your output is then its prbly not a common use of ai
    • applicative 4 hours ago
      Did you read even the title?
      • system2 3 hours ago
        Reread what he said maybe?
  • a11r 34 minutes ago
    Even a 4-bit quant of Qwen3.8 27b is indistinguishable from Gemini 3.7 flash in our internal tests. With an RTX5090 card and ninfer, you can get ~800 TPS token generation (c=8) and ~140 Tokens per second single stream.
  • walrus01 1 hour ago
    Much of this is why I stick to the rule of:

    a) Don't quantize your KV cache

    b) Don't run quantizations of the LLM that are worse than the best available Q8 (the largest possible file size unsloth GGUF for a given model like qwen 3.8 27B as an example). I would rather things go slowly but I have confidence that it's doing things more accurately.

  • InvertedRhodium 2 hours ago
    I’m running Qwen3.8 aggressive uncensored Q4_K_P on a 4090 in a loop against the 2026 CrackMe CTF challenges.

    Using oh-my-pi in a prebuilt environment that I let Qwen build too.

    Codex wouldn’t even look at the files - literally, as soon as it read something with CTF it shut down. Didn’t even offer to fall back to a dumber model.

    • CamperBob2 2 hours ago
      How's it performing on the challenges?
      • InvertedRhodium 1 hour ago
        I only kicked this off last night before bed, so I've just got up to see the result of the first task.

        Challenge: Wallpaper

        https://github.com/crackmesone/ctf-2026-challenges-public/tr...

            Duration:     4h 00m 15s
            Termination:  completed
            Verdict:      PARTIAL
            Confidence:   0.95
        
        I'm using Kimi K3 as the evaluator because, again, Codex and co. wouldn't even evaluate the output. Kimi's verdict:

        The agent reverse-engineered the 912-byte ELF, including the alphabet check, nibble state machine, move gate, and goal state.

        It eventually produced:

            CMO{10012232101230103012333221101033210010}
        
        I independently verified the underlying input against the actual binary:

            printf '10012232101230103012333221101033210010' |
                ./wallpaper/handout/wallpaper
        
        which returns:

            good job, validate with CMO{your_input}
        
        and exits 0.

        The wrinkle is that the official answer key is:

            CMO{1012321103210033011233322110103321001}
        
        So the puzzle apparently admits multiple accepted inputs. The agent found a valid password by reverse-engineering the program, but did not recover the canonical secret from the answer key.
        • huseyinkeles 40 minutes ago
          I don't know why but your post was marked as [dead] for some reason. Just vouched for it.
          • InvertedRhodium 38 minutes ago
            Thanks! Might have been the multiple attempts at getting it to format nicely in a short amount of time.
      • mlvljr 1 hour ago
        [dead]
  • JacobJack 3 hours ago
    > And the comparisons in this post are not going to be running some 2.58-bit-gguf-in-ollama with a couple test prompts.

    Genuine question : is there something fundamentally wrong with Ollama ?

    I use Ollama because it is easy to set up and manage (and also because VLLM is not super Windows friendly).

    I thought the main advantage of VLLM was better concurrency management (better batching).

    But if the quality of the interference itself is an issue, then maybe I should reconsider my choice.

    • embedding-shape 3 hours ago
      People who use Ollama generally (not everyone obviously) don't always clearly understand what quantization they use when running models, so people end up saying "I tried running Qwen 3.8 27b locally and it was dumb" while Ollama would default to a Q4 version of the model, which has very different results from the BF16 weights, doesn't really speak to the model itself because it's been so quantized in that case.

      Sure, makes things easier, but tons of people misunderstand what they're using, then base and share their experiences on that, without really specifying what exact weights they use too.

      For a single local user, using llama.cpp directly shouldn't be a problem if you're already using Ollama's CLI, it works basically the same except you manage weights yourself, and if you put your favorite agent to make sense of the faux "registry + image layers" Ollama has prepared locally for you, you can reuse the files you've already downloaded with Ollama.

    • kangalioo 3 hours ago
      From what I've heard, Ollama has a bad reputation because it's a thin wrapper around llama.cpp without attributing it properly, thereby stealing recognition from the maintainers doing most of the work
      • b112 3 hours ago
        It seems, and that seems is entirely my unvalidated impression, that Ollama lags in features, as they're integrating after the fact those changes. But (seriously) an LLM told me that, when some aspects of MoE models were better supported with the latest llama.

        And it did in that case make a significant difference.

        • smcleod 2 hours ago
          It's very far behind llama.cpp, vLLM and SGLang in features yes. In part because of that but also due to some poor default settings it generally performs a lot worse as well.
    • cube00 3 hours ago
      > is there something fundamentally wrong with Ollama ?

      Yes https://news.ycombinator.com/item?id=47788385

    • zepearl 2 hours ago
      I know nothing about "vLLM" -> anything better compared to "llama.cpp"?

      I started with "Ollama" (precompiled version) and it worked and was good enough to understand the very basics.

      Then I downloaded the sourcecode of "llama.cpp", compiled it with specific compilation options for my GPUs (CUDA/nVidia using proprietary module on Gentoo Linux) & CPU (AMD), and the same model ran twice as fast -> since then I stuck with "llama.cpp" (and "ik_llama.cpp" in very few cases).

      I honestly don't know what made "Ollama" (precompiled) so much slower than "llama.cpp" (compiled locally) at that time and I'm too lazy to doublecheck now, in any case I now absolutely love all the knobs that "llama.cpp" has to tune your hardware setup & your workload, which is the reason why I recommend it.

    • VCFundedGenYer 2 hours ago
      In my limited testing of ollama (in a Docker container because I don't want Meta touching bare metal disk), I noticed that first of all the model is abandoned. Last updated in 2023. Secondly, it basically can't do anything. Anything I asked of it resulted in a lot of "I can't do that" so it's just stupid.
      • Macha 2 hours ago
        Are you confusing llama (the model from meta that popularised a lot of open AI stuff but is quite dated now) and ollama (the set of tools for running models, originally llama but no longer limited to it, not maintained by meta)
  • catlifeonmars 3 hours ago
    > I will make you read the really long unpleasant version with math.

    This is the version I want to read :)

    I assume it is unpleasant in spite of the math, not because of it?

    • a1o 2 hours ago
      I thought it was a link too because of the line under the with math but it isn’t. :/
  • nineteen999 14 minutes ago
    Can we stop posting this bullshit please?

    Local models running on your GTX 5090 or Macbook M99 are never going to reach the levels of the currently hosted "frontier models" with hundreds/thousands of CPU's and GPU's. Get a grip.

    Anybody who's tested this over the last 12-18 months knows this. Don't waste our time because you're late to the party.

  • IronWolve 1 hour ago
    sglang, 150+ tok/s on a 5090 in ubuntu 26.04 via wsl. gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090, dspark, medium reasoning, 96k context.

    Using opencode and it built a old fashioned arcade vertical shooter with no issues.

    Images are ok'ish, just had grok create updated images, and it came out great.

  • paulyy_y 30 minutes ago
    Slop images littered throughout the blog post -> opinions irrelevant. Simple as.
  • anotherCodder 4 hours ago
    most of the time when a local model feels dumb its not the quant, its the chat template. a lot of gguf mints just drop the template from the metadata and the runtime silently falls back to chatml. model still talks fine so nobody notices, it just gets noticeably dumber. got burned by this myself serving qwen, now i grep the gguf for the template tokens before i blame anything else. second place is sampling, people run whatever defaults their ui ships instead of what the vendor recommends and then compare that to benchmark numbers that were run greedy or with the official settings
    • washadjeffmad 1 hour ago
      I've been comparing against TextGen and llama.cpp while I port to LocalAI and have been surprised by what's happening over the API, even with the defaults and jinja. It's been a fair reminder not to eschew familiarizing myself with the repos.
  • koba3 1 hour ago
    [dead]