Every checkpoint is trained, quantized, and validated on our own DGX B300 capacity in Canada — benchmarked in the open, and served through vLLM.
Frontier-tier open models — 600B+ parameter MoEs, long-context dense models — need real Blackwell silicon to train and serve at production scale. We operate it here.
What we've put on Hugging Face under open licenses — now past 56,000 downloads across ten builds — and what we're calibrating next. Plus an invitation: if there's a model your team needs quantized in the open, we'd like to hear about it.
Our own drafter, not a quant: a self-trained DFlash2 block-diffusion speculative-decoding drafter for GLM-5.3-Flash, trained against our W4A16-MTP checkpoint on 736,675 samples the target model generated itself — no third-party drafter weights or traces anywhere in the training path. The third release in the lineage (-E 15 Sep → -F 24 Sep → -G), and the first to beat the widely-used CC-BY-NC-ND reference drafter on acceptance, outside the noise band, on identical hardware. Drafter-of-record on our 2× DGX Spark stack since 25 Sep 2026, and the latency drafter in our 8× H100 recipe.
A 4-bit weight quant of Z.ai's GLM-5.3-Flash — a multimodal MoE with a 1M-token context — that keeps the BF16 MTP draft head and the BF16 vision tower intact. At 177.7 GiB (−70% vs BF16) it serves the full 1M-token context on 2× DGX Spark or 2× H200, and 262K–512K on 4- and 8-GPU H100, H200, and RTX PRO 6000 configurations. On 8× H100 it leads NVIDIA's own NVFP4 checkpoint on the same rig: +5.1% at c256 like-for-like, and +47.8% single-stream against NVFP4 run without speculative decoding. It also carries the first formal vision scores for this model. Only the 36,288 routed-expert GEMMs are INT4 (g=128, GPTQ); attention, router, shared experts, embeddings, the vision tower, and the MTP head ride at full precision.
A 4-bit weight quant of Tencent's Hy3 (295B MoE, 21B active) that keeps the multi-token-prediction draft layer in BF16 — matching the FP8 release on quality at 57% of its footprint, and taking the high-concurrency throughput crown among 4-bit Hy3 quants. Routed experts to INT4 (g=128, GPTQ); attention, dense layer, shared expert, router, embeddings, LM head, and the MTP head all in BF16.
A 4-bit weight quant of GLM-5.2 (744B MoE) that keeps the multi-token-prediction draft head in BF16 — matching the FP8 release on quality, fitting on four H200s instead of eight, and the fastest of the popular 4-bit GLM-5.2 quants in the interactive serving regime. Routed experts to INT4 (g=128, GPTQ); attention, dense prefix, shared experts, router, embeddings, LM head, and the MTP head all in BF16.
The first NVFP4-FP8 quant of DeepSeek-V4-Pro with the MTP draft head preserved for vLLM speculative decoding. A byte-deterministic conversion of V4-Pro's native MXFP4+FP8 source to NVFP4 on 8× B300 SXM6 — shipped alongside the upstream vLLM patch that makes the flashinfer_trtllm NVFP4 MoE backend load it, merged the same day.
A W4A16 + FP8_BLOCK quant of DeepSeek-V4-Flash that keeps the MTP draft head at BF16 — the first build on this base where speculative decoding survives end-to-end through GPTQ calibration, transformers save, and vLLM load on Hopper. The reproduction repo documents three upstream silent-drop bugs and the fixup pipeline that routes around them.
Datacenter-Blackwell NVFP4 build of DeepSeek-V4-Flash with the MTP draft head retained on-disk, so vLLM serves it with --speculative-config method=mtp end-to-end. Same quant math as the comparable peer; the structural difference is that MTP survives calibration instead of being stripped at load time.
A W4A16 + FP8_BLOCK quant of DeepSeek-V4-Flash with a vLLM serving recipe. Runs on Hopper (H200) and two Blackwell SKUs (DGX Spark, RTX PRO 6000), built alongside the upstream patches it depends on. Open weights, no token required.
Each lands on Hugging Face under the base model's license when it's ready. Status updates: [email protected].
We ship one or two community quants per cycle. We're looking for:
First-look review within two weeks. Selected requests get a public reproduction repo and a model card under canada-quant/.