Join the discussion
Write your take first — we'll ask for email only when you're ready to publish.
- Hacker News
- I suspect we're going to see a resurgence of local computing because of LLMs, although not in exactly the same way as before.by voxleone
- around this time, i was thinking about running a local LLM for the company i work for. i'm trying to figure out what kind of hardware i'd need and what level of performance i could expect. i tried a few things on your website but how much do the results reflect real world usage?
in my use case, the model might need to analyze a given task first then generate and execute a database query and finally send the results back to the client. to what extent could your setup handle this kind of workflow and how accurate are the results?
good job.
- Projects like Strata show that this sort of thing will become outdated very fast (unless they can start including estimations for stuff like that as well).
Increasingly, people are finding clever technical solutions to get the most out of whatever hardware you have.
by Grombobulous - What surprised me is how fast the Mac Studio is and how slow the Spark is. I thought they were at least comparable, but the former is getting 20 tok/s and the latter 4.6 tok/s with e.g. Qwen 3.8 27B and Q8.by sgt
- Prices for everything are all in flux at this moment. Consumer cards like RTX 5090 are bonkers. Even 3090s have like doubled in the last six months. Meanwhile, prices for renting this class of hardware on vast.ai are spiking too, with rates increasing like 50% in September alone.
It's hard to know if we're at the beginning or the end of this trend. If manufacturing continues to focus on high end datacenter gpus at $40k a pop, and we continue to have supply constraints, this could last a while.
On the other hand, if Chinese models continue to nip at the tails of the frontier, cheaper by 20x, running on new Chinese chips, the whole bottom could fall out of the US market, and things could get very weird.
by montroser - It says 46 tok/sec for Qwen3.6-27b 4 bit on RTX5090, but I get 60-70 tok/sec on a small context like that (5 bit weights with 8 bit KV cache), although it does slow down to 40 tok/sec on large contexts. If I add ngram speculation, then it can increase to 100+ tok/sec on code.
There are many knobs to tweak to make it faster, which the site ignores.
by kgeist - By running vllm at work I learned that you will usually outperform benchmarks which you will find on the internet. Mostly by running proper parameter tuning. So you get impressive numbers. Still in reality those numbers do not tell the whole truth. We found that real usage is more 60k-140k prompt with a few k of output for realistic agentic coding workloads. Most of the time that will give you half to a third of decode throughput vs the benchmark. It also becomes much more about cache aware routing and kv cache management (how to offload and restore). Because prefilling 60-140k will simply take too long for interactive work.by Jabdoa2
- I feel like these types of benchmarks mostly test what the mainstream tells you is possible and or works. You rarely get to hear the different implementations of inference engines, different types of setups (e.g. hybrid ones or older hardware) and their pros and cons. Also it seems these were tested with very minimal token usage.
Currently on my 9 year old laptop with a GTX 1070 Max-Q 8GB vram, 16 gb ddr4 ram and i7 8750h. I am running a finetuned Qwen 3.5-35B-A3B (Tiel-Coder-35B-A3B-GGUF-MTP) with 4-bit quant on a llamacpp fork (Turboquant) with 131k context at around 16-19 tk/s and 350-500 tk/s prefil using Pi harness. It is amazing for writing scripts and doing medium coding tasks and summarizing text for me. After around 100k context it goes down to 9 tk/s but that is to be expected.
If you ever ask any frontier LLM or these types of benchmarks, they would probably say that the best model I can run on my system is a crappy 9B model or something, which is just completely false. So I really wonder how much more you can pull out from more powerful systems and much better open weight models.
by lopotubez