Nestmodels for Linnet
Models / qwen3-4b

Qwen3 4B

A 4B-parameter Qwen3 decoder with grouped-query attention, per-head RMS normalization of the query and key projections, unbiased projections, and an output head tied to the token embedding.

4B parametersqwen3Apache-2.0text-generationdecoder-onlygrouped-query-attentionqk-norm
Hugging FaceGitHubarXiv Registry entry

Greedy decoding of one chat prompt, through Linnet and through transformers (KV cache, argmax, bf16).

Prompt

Explain in two sentences why the sky is blue.

Linnet

<think> Okay, the user wants a two-sentence explanation of why the sky is blue. Let me start by recalling the basic concept. The sky appears blue because of Rayleigh scattering. When sunlight enters the Earth's atmosphere, the shorter wavelengths of light, like blue, scatter more than the longer wavelengths, such as red. That's why we see blue in the sky. Wait, but I need to make sure I'm not mixing up any details. Rayleigh scattering is

transformers (KV cache, argmax, bf16)

<think> Okay, the user wants a two-sentence explanation of why the sky is blue. Let me start by recalling the basic concept. The sky appears blue because of Rayleigh scattering. When sunlight enters the Earth's atmosphere, the shorter wavelengths of light, like blue, are scattered more by the gas molecules. That means blue light is spread out in all directions, making the sky look blue. But wait, I need to make sure I'm not mixing up any details.

The first 55 of 96 generated tokens agree with the reference; after that the two stacks round a near-tie differently and part ways, which bf16 greedy decoding allows.