Qwen3-4B — BRIK (int4)
Qwen3-4B (Alibaba, Apache-2.0) converted to the BRIK format so it runs inside the browser on the visitor's GPU (WebGPU) — no inference server, no API key.
The heavy desktop option of the Brimkern catalogue: the largest model the engine streams today, for machines with the VRAM to hold it.
| File | qwen3-4b-q4.brik (2.53 GB) |
| Quantization | int4, group of 32 — pre-quantized in the layout the kernels read (no dequantization at load) |
| Tokenizer | embedded in the file (no separate download) |
| Architecture | llama-family with QK-norm — 36 layers, d=2560, vocab 151 936 |
| Engine | Brimkern — hand-written WGSL kernels |
Try it in one click
👉 https://brimkern.com/chat?model=romainkh14/Qwen3-4B_BRIK
⚠️ 2.53 GB streamed on first use, and roughly the same again in VRAM. The page announces the cost before the download starts, the transfer resumes after a drop, and the file is reusable offline afterwards. On a laptop GPU expect a first load in minutes, not seconds.
Measured
No per-model throughput figure is published here yet: our decode benches cover the small models people load by default (see LFM2.5-230M and Qwen2.5-0.5B). What is measured on this class of model is that decoding is memory-bandwidth-bound, so throughput follows the GPU's bandwidth almost linearly — the engine sits at 79-92 % of the machine's measured ceiling.
The rule in this project is that a number gets published only if a replayable bench produced it. This one has not been run on 4B yet.
Format
A .brik is a self-describing container: architecture, tokenizer and configuration travel inside
the file, weights are pre-quantized in the layout the kernels read, and every layer is one contiguous
HTTP range — which is what makes a 2.53 GB model loadable in a tab at all. Specification:
BRIK_FORMAT.md.
License
Weights under Apache-2.0 (Qwen3-4B, Alibaba). The Brimkern engine itself is MIT.