Qwen3-4B-Thinking-2507 (opens in new tab)

https://huggingface.co/lmstudio-community/Qwen3-4B-Thinking-...

KV cache size = 2 * batch_size * context_len * num_key_value_heads * head_dim * num_layers * element_size. The "2" is for the two parts, key and value. Element size is the precision in bytes. This model uses grouped query attention, which reduces num_key_value_heads compared to a multi head attention (MHA) model.

With batch size 1 (for low-latency single-user inference), 32k context (recommended in the model card), fp16 precision:

2 * 1 * 32768 * 8 * 128 * 36 * 2 = 4.5GiB.

I think, anyway. It's hard to keep up with this stuff. :)

For a while it was china vs. world, but lately it's even more divided, with heavy camping on specific models. You can still get some signal, but you have to either ban a lot of accounts, or read new during different tzs so you can get some of that "i'm just here for the tech stack" vibe from posters.

The new qwen3 model is not out yet.

It's an interesting proxy, but idk how reliable it'd be.

I am running this beast on my dumb pc with no gpu, now we are talking!

Claude is not cheap, why is it far and away the most popular if it's not top 10 in performance?

Qwen3 235b ranks highest on these benchmarks among open models, but I have never met someone who prefers its output over Deepseek R1. It's extremely wordy and often gets caught in thought loops.

My interpretation is that the models at the top of ArtificialAnalysis are focusing the most on public benchmarks in their training. Note I am not saying XAI is necessarily nefariously doing this, could just be that they decided it's better bang for the buck to rely on public benchmarks than to try to focus on building their own evaluation systems.

But Grok is not very good compared to the anthropic, openai, or google models despite ranking so highly in benchmarks.