Qwen3

The 1.7B base checkpoint on Apple Silicon. Qwen3 is not Qwen2.5 with a new number on it — the attention block gained QK-Norm and lost the qkv bias — so it needed its own adapter and its own proof that the adapter is right.

What it costs on the device

Measured on an Apple M4 Pro, 2026-09-20, with the published core. 0 MiB copied verbatim.

The command

$ huggingface-cli download Qwen/Qwen3-1.7B

$ veizik run 1.7b "Explain attention."
  device      : Apple M4 Pro
  model       : qwen3-1.7b  H 2048 layers 28
  checkpoint  : 3281 MiB checkpoint
  weights     : 923 MiB (0 MiB copied verbatim)

How the adapter was proved

A new architecture is where quiet wrongness lives. The check was not "does it produce text" but "does it produce the same numbers as an implementation we already trust".

Architecture support

qwen3 — EXPERIMENTAL. Import, on-device preparation and tokenizer validated.

Reference used

A numpy/torch reference implementation, itself validated against transformers' Qwen3ForCausalLM.

Numeric result

Max absolute logit delta 4.3e-5; argmax identical at every position.

What differs from qwen2

Adds QK-Norm (self_attn.{q,k}_norm.weight, 128) and drops the qkv bias.

Context

Opened to 32,768 on 2026-09-16 and verified by recall.

Quality

Uncalibrated quantisation, as with the rest of this release.

How this family got here