1-bit 27B in the browser: 25–30 tok/s on a 6 GB RTX 3060 Laptop (WebGPU, no install)

mentria.ai is a browser inference engine I've been building solo, from scratch in WebGPU/WGSL. This week it crossed a milestone I had been chasing for a while: a 27B one-bit model answering at up to 30 tokens/s on an RTX 3060 Laptop GPU with 6 GB of VRAM, in Chrome, from a web page. No install, no server, nothing leaves the machine. The model. Bonsai-27B, a natively 1-bit model trained and released by Prism ML. I repacked it for our engine, wrote the kernels that run it, and checked its quality myself (full eval deltas vs FP16 Qwen3.6-27B are in the HF card). Every weight is one sign bit with one scale per 128 weights, about 1.14 bits per parameter, so 27 billion parameters sit in 3.8 GB of GPU memory. Repack and numbers: huggingface.co/mentriaai/Bonsai-27B-mentria The engineering that got it to 30. Two days before this post the same model decoded at 15 tok/s on this laptop. Decode is memory-bound: one word of the 27B is 804 GPU dispatches, and 401 of them are the 1-bit matrix-by-vector kernel that streams the model's 3.6 GB of matmul weights once per word (embeddings bring the file to 3.8 GB). The kernel that won was the one written for phones: four 1-bit weights have only 16 possible partial answers, so it computes all 16 once into on-chip scratch and each row reads its answer from that table instead of multiplying (station 258 — the numbered write-ups on the facts page). On Ampere that table hit a wall that was not the weights at all: scratch memory has 32 banks, the table's addressing sent sixteen threads of every warp to the same bank, and the kernel ran at 38% of the card's bandwidth. One padding slot per row spread the addresses across the banks; the kernel's share of a word dropped from 26.5 ms to about 21, and on top of the 15 → 26 tok/s the table itself had bought, the laptop hit 32 tok/s raw — the 25–30 is what survives sampling and UI overhead in the chat (s259). The prompt stage had the same disease in its tile layout; a retile took a 1,489-token prompt from 29.6 s to 25.3 s. Every change shipped only after its output was byte-identical to the build before it, which is also why the kernel adds its partial sums in a fixed order and accepts a 44% occupancy ceiling. Every claim here has a numbered write-up on the engine facts page. Numbers (RTX 3060 Laptop 6 GB, Windows 11, Chrome 152 on D3D12, fresh loads): Decode: 25–30 tok/s in the chat UI once the card is warm. Prompt processing: a 1,489-token prompt in about 25 s. Context: 3,072 tokens on this 6 GB card; 8,192 on 16 GB Macs; more on bigger cards, at 128 KiB per token. The KV cache is exact math, no quantized cache. The next step on 6 GB is consolidating the engine's few thousand small GPU buffers into a handful of large arenas, so the driver stops holding about 300 MiB of slab slack; that is the arithmetic for 4,096, and it is not built yet. Load: under 10 s from the browser cache; the first download is 3.8 GB, once. Also in the engine: the smaller tiers (Qwen3.5 0.8B, 2B, 4B) cover phones and low end devices, LoRA adapters of a few MB hot-swap at the matmul in under a second, and there is a vision tower for image input. Try it at https://mentria.ai/tools/ai-chat/ (the 27B tier appears when your GPU qualifies). The code that ships, the benchmarks and how I measure are in the repo: https://github.com/mentria-ai/website. The engine facts page, for how the engine and the model actually work: https://mentria.ai/assets/learn/engine-facts.html