[HN Gopher] Show HN: LocalScore - Local LLM Benchmark
       ___________________________________________________________________
        
       Show HN: LocalScore - Local LLM Benchmark
        
       Hey Folks!  I've been building an open source benchmark for
       measuring local LLM performance on your own hardware. The
       benchmarking tool is a CLI written on top of Llamafile to allow for
       portability across different hardware setups and operating systems.
       The website is a database of results from the benchmark, allowing
       you to explore the performance of different models and hardware
       configurations.  Please give it a try! Any feedback and
       contribution is much appreciated. I'd love for this to serve as a
       helpful resource for the local AI community.  For more check out: -
       Website: https://localscore.ai - Demo video:
       https://youtu.be/De6pA1bQsHU - Blog post:
       https://localscore.ai/blog - CLI Github:
       https://github.com/Mozilla-Ocho/llamafile/tree/main/localsco... -
       Website Github: https://github.com/cjpais/localscore
        
       Author : sipjca
       Score  : 117 points
       Date   : 2025-04-03 16:32 UTC (4 days ago)
        
 (HTM) web link (www.localscore.ai)
 (TXT) w3m dump (www.localscore.ai)
        
       | jborichevskiy wrote:
       | Congrats on launching!
       | 
       | Stoked to have this dataset out in the open. I submitted a bunch
       | of tests for some models I'm experimenting with on my M4 Pro.
       | Rather paltry scores compared to having a dedicated GPU but I'm
       | excited that running a 24B model locally is actually feasible at
       | this point.
        
       | mentalgear wrote:
       | Congrats on the effort - the local-first / private space needs
       | more performant AI, and AI in general needs more comparable and
       | trustworthy benchmarks.
       | 
       | Notes: - Olama integration would be nice - Is there an anonymous
       | federated score sharing? That way, users you approximate a
       | model's performance before downloading it.
        
         | sipjca wrote:
         | Can you tell me more about the "anonymous federated score
         | sharing"? Maybe something we can think about more
         | 
         | I totally agree with Ollama integration and if there is
         | interest we will try to upstream into llama.cpp
        
       | alchemist1e9 wrote:
       | Really awesome project!
       | 
       | Clicking on GPU is a nice simple visualization. I was thinking
       | maybe try to put that type of visual representation intuitively
       | accessible immediately on the landing page.
       | 
       | cpubenchmark.net could he an example technique of drawing the
       | site visitor into the paradigm.
        
         | sipjca wrote:
         | I think you might be right, definitely interested in this
         | feedback and creating charts and graphs that are the most
         | useful for folks!
        
       | roxolotl wrote:
       | This is super cool. I finally just upgraded my desktop and one
       | thing I'm curious to do with it is run local models. Of course
       | the ram is late so I've been googling trying to get an idea of
       | what I could expect and there's not much out there to compare to
       | unless you're running state of the art stuff.
       | 
       | I'll make sure to run contribute my benchmark to this once my ram
       | comes in.
        
       | jsatok wrote:
       | Contributed scores for the M3 Ultra 512 GB unified memory:
       | https://www.localscore.ai/accelerator/404
       | 
       | Happy to test larger models that utilize the memory capacity if
       | helpful.
        
         | deanputney wrote:
         | That's very interesting. I guess it just can't compete with any
         | of the Nvidia cards? I would think your results should show up
         | if sorted by "generation"- maybe the leaderboard is cached...
        
           | sipjca wrote:
           | Ty for pointing this out, the results are taken from the db
           | based on LocalScore, I will make some modifications to make
           | the sorting better here
        
           | zamadatix wrote:
           | Non-VRAM based options of the moment are only competitive
           | when it comes to the size of a model they can run, since the
           | alternative is "you can't run it at that price otherwise".
           | 
           | Bandwidth at a given size is king, only then followed by
           | enough compute to utilize it.
        
       | ftbsqcfjm wrote:
       | Interesting approach to making local recommendations more
       | personalized and relevant. I'm curious about the cold start
       | problem for new users and how the platform handles privacy.
       | Partnering with local businesses to augment data could be a smart
       | move. Will be watching to see how this develops!
        
       | omneity wrote:
       | This is great, congrats for launching!
       | 
       | A couple of ideas .. I would like to benchmark a remote headless
       | server, as well as different methods to run the LLM (vllm vs tgi
       | vs llama.cpp ...) on my local machine, and in this case llamafile
       | is quite limiting. Connecting over an OpenAI-like API instead
       | would be great!
        
         | sipjca wrote:
         | LocalScore dev here
         | 
         | Thank you! I think this is quite possible! If you don't mind
         | starting a discussion on this I would love to think aloud there
         | 
         | https://github.com/cjpais/LocalScore/discussions
        
       | sharmasachin98 wrote:
       | This looks super useful, especially with so many folks
       | experimenting with local LLMs now. Curious how well it handles
       | edge devices. Will give it a try!
        
       | gunalx wrote:
       | Why choose a combibation of llama and qwen, when you could have
       | used just qwen models with more permissive license?
        
         | sipjca wrote:
         | It's kind of just an artifact of the development. Happy to
         | switch the default models and sizes in the future especially
         | based on community feedback
        
       | FloatArtifact wrote:
       | I've been waiting for something like this. Have you considered
       | the following based on the benchmark data that's submitted beyond
       | the GPU?
       | 
       | 1. User selects a model, size and token output speed and latency.
       | The website generates a hardware list of components that should
       | match the performance requirements.
       | 
       | 2. User selects hardware components and the website generates a
       | list of models that performant on that hardware.
       | 
       | 3. Monetize through affiliate links the components to fund the
       | project. Think like PC part picker.
       | 
       | I know there's going to be some variability in the benchmarks due
       | to the software stack, but it should give a AI enthusiasts, an
       | educated perspective on what hardware can be relevant for their
       | use case.
        
         | sipjca wrote:
         | Right now the main priority is just getting the data out, but
         | in the future may have some interest in this. Or perhaps we can
         | open an API for others to build this as well
        
       | chrismorgan wrote:
       | I'm curious: does this fundamentally _need_ to contain an actual
       | model, or would it be okay if it generated a synthetic model
       | itself, full of random weights? I'm picturing downloading just,
       | say, a 20MB file instead of the multi-gigabyte one, and...
       | 
       | Hang on, why is https://blob.localscore.ai/localscore-0.9.2
       | 380MB? I remember llamafile being only a few megabytes. From
       | https://github.com/Mozilla-Ocho/llamafile/releases, looks like it
       | steadily grew from adding support for GPUs on more platforms, up
       | to 28.5MiB1 in 0.8.12, and then rocketed up to 230MiB in 0.8.13:
       | 
       | > _The llamafile executable size is increased from 30mb to 200mb
       | by this release. This is caused byhttps://github.com/ggml-
       | org/llama.cpp/issues/7156. We're already employing some
       | workarounds to minimize the impact of upstream development
       | contributions on binary size, and we're aiming to find more in
       | the near future._
       | 
       | Ah, of course, _CUDA_. Honestly I might be more surprised that
       | it's _only_ this big. That monstrosity will happily consume a
       | dozen gigabytes of disk space.
       | 
       | llamafile-0.9.0 was still 231MiB, then llamafile-0.9.1 was
       | 391MiB, now llamafile-0.9.2 is 293MiB. Fluctuating all over the
       | place, but growing a lot. And localscore-0.9.2 is 363MiB. Why
       | 70MiB extra on top of llamafile-0.9.2? I'm curious, but not
       | curious enough to investigate concretely.
       | 
       | Well, this became a grumble about bloat, but I'd still like to
       | know whether it would be feasible to ship a smaller localscore
       | that would synthesise a suitable model, according to the size
       | required, at runtime.
       | 
       | --***--
       | 
       | 1 Eww, GitHub is using the "MB" suffix for its file sizes, but
       | they're actually mebibytes (220 bytes, 1048576 bytes, MiB). I
       | thought we'd basically settled on returning the M/mega- prefix to
       | SI with its traditional 106 definition, at least for file sizes,
       | ten or fifteen years ago.
        
         | sipjca wrote:
         | LocalScore dev here
         | 
         | Llamafile could certainly be released without the GPU binaries
         | included by default and it would slim down the size
         | tremendously.
         | 
         | The extra 70MiB is that the CUDA binaries for LocalScore are
         | built with CuBLAS and for more generations of NVIDIA
         | architectures (sm60->sm120), whereas Llamafile is built with
         | TinyBLAS and for just a few generations in particular
         | 
         | I think it's possible to randomize weights with a standard set
         | of layers, and maybe a possibility for the future
        
       | TheFlyingPanda wrote:
       | Awesome stuff, congrats on launching!
        
       | zamadatix wrote:
       | The run and/or troubleshooting steps for Windows should probably
       | include the note you need to install
       | https://developer.nvidia.com/cuda-downloads?target_os=Window...
       | if you have an Nvidia GPU (and probably something similar if you
       | have an AMD GPU?). As it is right now the steps happily get you
       | benchmarking your CPU and I'd say that might even be worth adding
       | a "Warning: The benchmark is operating in CPU only mode, press y
       | to continue if this is intended" type message to the program.
       | 
       | Edit: And for the same prompt and generated token counts it runs
       | ~4x slower than `ollama run hf.co/bartowski/Qwen2.5-14B-Instruct-
       | GGUF:Q4_K_M --verbose`. It's possible I'm mixing up a few things
       | there but my results also post in the same ballpark slower than
       | others with the same GPU so it seems something is up with the
       | application in either case.
        
       | david_draco wrote:
       | I don't know if I should trust and run this code. If it was
       | associated to Mozilla I would. It says it is a Mozilla Builders
       | project, but https://builders.mozilla.org/projects/ does not list
       | it. I don't see a way to verify that localscore.ai is associated
       | with Mozilla.
        
         | zamadatix wrote:
         | It doesn't seem they update the site often (the last 'latest'
         | post is from December) but they reposted something claiming the
         | same on X https://x.com/llamafile/status/1907917417118105751
        
       ___________________________________________________________________
       (page generated 2025-04-07 23:02 UTC)