Mkengin

There Seen to be more and more harness benchmarks out there, pretty interesting read:

https://neuralnoise.com/2026/harness-bench-wip/

I don't think so, I use GrapheneOS and I think I can't even use the USB-C port for anything other than charging (which should be configurable).

No, I am not affiliated with the website, I just want to see more discussions based on uncontaminated benchmarks and feel that people rely too much on benchmarks that companies can conduct themselves. If that is the case, I don't feel I can trust them. For general LLM capabilities, for example, I would also tend to rely on dubesor [1] rather than artificial analysis or similar leaderboards.

[1] https://dubesor.de/benchtable

I would assume that if a tool is there and the alternative too costly that they would use the tool instead of buring their project. Just today I stumbled over this for example, where they use GenAI as well: https://reddit.com/comments/1prqfsu

Not for coding, but today I stumbled upon these two building their passion project using GenAI, which would otherwise perhaps not be possible: https://reddit.com/comments/1prqfsu

It doesn't have to be hyped to be used, for example today I found these two building their passion project using GenAI, which would otherwise maybe not possible, who knows: https://reddit.com/comments/1prqfsu

This is just one example, but today I found this where two people build their passion project using GenAI for image generation (+ photoshop), maybe otherwise this project wouldn't even be possible: https://reddit.com/comments/1prqfsu

Though this Codex version isnt on the leaderboard, GPT-5.2-Medium already seems to be a bit better than Opus 4.5: https://swe-rebench.com/

At least on swe-rebench it does pretty well: https://swe-rebench.com/

Your experience seems to match the recent results from swe-rebench: https://swe-rebench.com/

According to SWE-Rebench Anthropic and OpenAI are really close in performance, while GPT-5.2 costs less than half the cost of CC per problem.

https://swe-rebench.com/

Interesting. So similar to the vision encoder + projector in VLMs?

I am eagerly awaiting swe-rebench results for November with all the new models: https://swe-rebench.com/

I like this one: https://swe-rebench.com/

Or use RL to beat any AI detectors: https://reddit.com/r/LocalLLaMA/comments/1lnrd1t/you_can_jus...

https://arxiv.org/abs/2311.13600

https://arxiv.org/abs/2410.22911

https://arxiv.org/abs/2409.16167

Thank you for testing, I will test GPT-OSS for my use case as well. If you're interested I have 8 GB VRAM, 32 GB RAM and get around 21 token/s with tensor offloading, I would assume that your setup should be even faster than mine with the optimizations. I use the IQ4_KSS quant (by ubergarm on hf) with ik_llama.cpp with this command:

$env:LLAMA_SET_ROWS = "1"; ./llama-server -c 140000 -m D:\ik_llama.cpp\build\bin\Release\models\Qwen3-Coder-30B-A3B-Instruct-IQ4_KSS.gguf -ngl 999 --flash-attn -ctk q8_0 -ctv q8_0 -ot "blk\.(19|2[0-9]|3[0-9]|4[0-7])\.ffn_.*_exps\.=CPU" --temp 0.7 --top-p 0.8 --top-k 20 --repeat_penalty 1.05 --threads 8

In my case I offload layers 19-47, maybe you would just have to offload 37-47, so "blk\.(3[7-9]|4[0-7])\.ffn_.*_exps\.=CPU"

Why Qwen2.5 and not Qwen3-30B-A3B-Thinking-2507 or Qwen3-Coder-30B-A3B-Instruct?