Canada Quant Labs

@canadaquant

Canada's open-weight model lab. Training, quantizing, and shipping sovereign models on B300 silicon. Built in Victoria, BC. cql.ca

Vancouver

Canada Quant Labs
@canadaquant
Oct 1
Excited to soon share the first Canadian Legal Benchmark - canlegalbench.com - open dataset and papers to follow.
4
Canada Quant Labs
@canadaquant
Sep 29
Your H100s are leaving 47% of GLM-5.3-Flash on the table. Our W4A16-FP8-MTP: 181 tok/s single-stream. NVIDIA's NVFP4 on the same rig: 123 tok/s— Same quality:
canada-quant/GLM-5.3-Flash-W4A16-MTP · Hugging Face
From huggingface.co
Canada Quant Labs
@canadaquant
Sep 19
Anthropic: We need to slow down due to security concerns. Also Anthropic: On September 25, new Claude Code sessions for users in your organization will start in auto mode by default.
1
Canada Quant Labs
@canadaquant
Sep 2
Local Eh-i Reading the model card explains the model card
canada-quant/GLM-5.3-Flash-W4A16-MTP · Hugging Face
From huggingface.co
Canada Quant Labs
@canadaquant
Jul 17
Released: Hy3 W4A16-MTP 🚀 A 4-bit quant of the 295B Hy3 MoE that keeps the Multi-Token-Prediction layer in BF16 — so speculative decoding still works. ~590GB → ~172GB. Runs on 4×H100. Negligible quality loss. Most 4-bit MoE quants drop MTP. We didn't.
canada-quant/hy3-w4a16-mtp · Hugging Face
From huggingface.co
1

Canada Quant Labs’s posts