Are you running qwen3.6-27b on one 3090 with your KV cache at q4? Ime there is significant long-context recall accuracy degradation at that precision. I prefer putting the KV cache at q8 and working with the 120k context
> Also, note that there's zero CUDA dependency. It runs entirely on Huawei chips.
That is a huge claim to make with no evidence.
I researched what you said, and I have found no statement to that effect in their paper[0], on huggingface[1], twitter[2], WeChat[3], or in their news release[4].
They only mention as a footnote in only the Chinese version of their news release that they plan to reduce inference costs with the Ascend 950 supernode when it releases[5]. The only mention of Huawei in their paper is that they validated a technique to lower interconnect bandwidth on Ascend NPUs and Nvidia GPUs[6].
> When it comes to information transfer and processing, light can do things that electricity can’t. Photons — particles of light — are far zippier than electrons at working their way through circuits.
Electrons themselves don't move at the speed of light, but information transfer (i.e. communication) via electrons does happen close to the speed of light.
A subtle, but important, distinction that's often misunderstood and means computational performance gains would probably come from bandwidth, not latency.
If OP meant they have the fastest implementation of Gemma 4 on Blackwell at the moment, I guess that is technically true. I doubt that will hold up when TensorRT-LLM finishes their implementation though.
I know Arc AGI 2 has a private test set and they have a good amount of results[0] but it's not a conventional benchmark.
Looking around, SWE Rebench seems to have decent protection against training data leaks[1]. Kagi has one that is fully private[2]. One on HuggingFace that claims to be fully private[3]. SimpleBench[4]. HLE has a private test set apparently[5]. LiveBench[6]. Scale has some private benchmarks but not a lot of models tested[7]. vals.ai[8]. FrontierMath[9]. Terminal Bench Pro[10]. AA-Omniscience[11].
So I guess we do have some decent private benchmarks out there.
It's easy to game and human evaluation data has its trade-offs, but it's way easier to fake public benchmark results. I wish we had a source of high quality private benchmark results across a vast number of models like Lmarena. Having high quality human evaluation data would be a plus too.
It's referring to the Lmsys Leaderboard/Lmarena/Arena.ai[0]. It's very well-known in the LLM community for being one of the few sources of human evaluation data.
Public benchmarks can be trivially faked. Lmarena is a bit harder to fake and is human-evaluated.
I agree it's misleading for them to hyper-focus on one metric, but public benchmarks are far from the only thing that matters. I place more weight on Lmarena scores and private benchmarks.
