CQL · MODELS & INFRASTRUCTURE

The model
catalogue.

Every checkpoint is trained, quantized, and validated on our own DGX B300 capacity in Canada — benchmarked in the open, and served through vLLM.

Shipped
6 open checkpoints
Formats
W4A16 · NVFP4 · MTP
Serving
vLLM · Hugging Face
I   /   CAPABILITIES

Hardware Canada
doesn't have yet.

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.

Hardware

  • PLATFORMNVIDIA DGX B300 · 8× Blackwell
  • MEMORY2.3 TB HBM3e total, NVLink Switch fabric
  • PRECISIONNative FP4 compute for both training and inference
  • LOCATIONHosted at Equinix Vancouver · sovereign by default
  • THROUGHPUT10–50× cloud-API throughput per dollar at steady state

Toolchain

  • BASECohere Command A+ · DeepSeek V4 · GLM-5.2 · Llama · Qwen — open-weight foundations
  • TRAININGTRL · Axolotl · DeepSpeed · custom GRPO trainer
  • QUANTIZATIONllm-compressor · GPTQ · AWQ · SmoothQuant · QuaRot
  • INFERENCEvLLM · FlashAttention-3 · CUTLASS kernels
  • EVALUATIONlm-eval-harness · custom domain suites
  • UPSTREAMActive contributors to vLLM and llm-compressor
II   /   MODELS

Open weights,
shipped and shipping.

What we've put on Hugging Face under open licenses — now past 56,000 downloads across six 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.

SHIPPED · 16 JUL 2026 · APACHE-2.0

Hy3
W4A16-MTP.

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.

172 GB
disk footprint · 57% of the FP8 release · 3.4× under BF16
144
tok/s @ c=1 with MTP k=2 · ties FP8 (143) · +40% vs the MTP-less GPTQ quant
2,146
tok/s @ c=32 · +21% over FP8 · fastest 4-bit Hy3 quant at concurrency
A-2.0
open weights · Apache-2.0, same license as the base
Serving footprint
~172 GB · down from ~598 GB BF16 and ~300 GB FP8 · 8×H100-80GB with headroom (4×H100 at TP=4)
Throughput vs FP8
ties FP8 @ c=1 (144 vs 143 tok/s, MTP k=2) · +21% @ c=32 (2,146 vs 1,854 tok/s)
vs the 4-bit field
+40% @ c=1 vs the MTP-less AngelSlim GPTQ quant (103 tok/s) · AWQ-INT4 edges c=1 at 152, trails at c=32 (2,028)
GSM8K / HumanEval
93.6% / 94.5% · FP8: 93.0% / 94.5%
Reasoning mode
AIME combined 76.7 · MATH-500 94.0 · GPQA-diamond 88.4 · FP8: 80.0 / 93.4 / 89.9 · all within noise
MTP speculative decoding
draft layer re-injected at BF16 post-quantization · 62.1% aggregate acceptance · ~2.24 tokens per step
Architecture
295B total, 21B active · MoE, 192 experts top-8 · W4A16 GPTQ routed experts (g=128) + BF16 attention, dense layer, shared expert, router & embeddings · MTP draft layer retained at BF16 · native 256K context
SHIPPED · 28 JUN 2026 · MIT

GLM-5.2
W4A16-MTP.

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.

4×H200
serving footprint · half the GPUs vs the 8×H200 BF16 release
132
tok/s @ c=1 · fastest 4-bit GLM-5.2 quant (+69% vs next)
82.0%
SWE-bench Verified · FP8 is 82.2%
MIT
open weights · same license as the base
Serving footprint
~405 GB on 4×H200 · down from ~1.49 TB BF16 on 8×H200
Throughput vs FP8
+48% @ c=1 · +32% @ c=8 (where MTP helps most) · ~13% slower at full saturation
vs the 4-bit field @ c=1
132 tok/s vs 78 AWQ-INT4 · 74 NVFP4 · same harness, output tok/s
GSM8K / MATH-500
0.960 / 0.954 strict · FP8: 0.955 / 0.958
IFEval prompt / inst strict
0.909 / 0.911 · FP8: 0.891 / 0.903
Long-context retrieval
RULER @ 32K / 64K: 0.832 / 0.841 · full 1M context on 8×H200 · up to 128K on 4×H200
Architecture
744B total · MoE · W4A16 GPTQ routed experts (g=128) + BF16 attention, dense prefix, shared experts, router & embeddings · MTP draft head retained at BF16
SHIPPED · 22 MAY 2026 · MIT

DeepSeek-V4-Pro
NVFP4-FP8-MTP.

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.

+36.9%
batched throughput vs upstream MXFP4 · c=16
96.59%
GSM8K · 1,274/1,319, 0 truncations
#42209
vLLM NVFP4 MoE PR · merged 22 May
MIT
open weights · same license as the base
Batched throughput
672.6 tok/s vs 491.4 upstream MXFP4 @ c=16 · 8× B300 SXM6 TP=8 + EP
Single-stream + MTP
139.3 tok/s with MTP n=1 vs 110.8 baseline · +25.7% @ c=1
GSM8K
96.59% strict-match (1,274/1,319) · 0 truncations
AIME 2024 pass@1
70.0% (21/30) · Non-Think greedy · 0 truncations
MTP draft acceptance
91.21% vs 90.92% MXFP4 on focused probe · 92.83% cumulative (40,225 drafts)
Architecture
1,598.84B total · 49.60B active per token · MoE · NVFP4 experts (g=16, E4M3) + FP8_BLOCK (128×128) attention & shared experts · MTP retained byte-identical to native
SHIPPED · 22 MAY 2026 · MIT

DeepSeek-V4-Flash
W4A16-FP8-MTP.

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.

1.49×
decode speedup with spec-decode · bs=1, k=1
22.3K
Hugging Face downloads · most-downloaded build in the org
93.71%
GSM8K 8-shot strict · ± 0.67
MIT
open weights · same license as the base
Decode TPOT
6.02 ms with spec-decode vs 8.93 ms without · 8× H200 SXM5 TP=2
GSM8K 8-shot
93.71% ± 0.67 strict-match
MMLU 5-shot
86.88% ± 0.27
HumanEval pass@1
84.76% ± 2.82
ToolCall15
24 / 30 (80%)
Architecture
284B total · 13B active per token · MoE · W4A16 GPTQ experts (g=128) + FP8_BLOCK attention · MTP retained at BF16 · 159 GB / 4 shards
SHIPPED · MAY 2026 · MIT

DeepSeek-V4-Flash
NVFP4-FP8-MTP.

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.

3.5×
compression: ~600 GB BF16 → 172 GB
96.0%
AIME 2024 non-truncated · matches BF16+MTP
88%
MTP draft acceptance · flat c=1 → c=16
MIT
open weights · same license as the base
AIME 2024 pass@1
96.0% non-truncated · 83.3% raw (65K cap) · 4× B300 SXM6 TP=4
GSM8K 8-shot
91.81% strict-match · 95.15% flexible-extract
HumanEval pass@1
91.5% EvalPlus · 84.8% EvalPlus+
IFEval prompt-strict
85.4%
Coding throughput
278.7 tok/s @ c=1 · 1,577 tok/s @ c=16 · 4× B300 TP=4
Architecture
284B total · 13B active per token · MoE · NVFP4 experts (g=16, E4M3) + FP8_BLOCK (128×128) attention · MTP retained at BF16 · 172 GB / 35 shards
SHIPPED · MAY 2026 · MIT

DeepSeek-V4-Flash
W4A16-FP8.

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.

19.6K
Hugging Face downloads · all-time
3.8×
compression: 543 GB → 143 GB
500K
token context · NIAH-validated single-session
MIT
open weights · same license as the base
GSM8K 8-shot
95.45% Spark TP=2 · 95.07% RTX PRO 6000
HumanEval pass@1
80.49% Spark TP=2 · 78.05% RTX PRO 6000
Toolcall15
92% Spark · 90% RTX PRO 6000
Long-context NIAH
PASS — 256K × 2 concurrent · 500K × 1
Decode throughput
84 tok/s output @ bs=2 · 2× RTX PRO 6000
Architecture
284B total · 13B active per token · MoE · W4A16 experts + FP8_BLOCK attention
IN FLIGHT
Planning
Reasoning & agentic distill
Distilling frontier reasoning and agentic coding into a deployable open-weight quant.
Prototyping
Autonomous coding cluster
Cluster-scale autonomous coding framework across our Blackwell fabric.
Designing
CanLegal-Bench
Open eval suite for Canadian common-law and civil-code reasoning, with citation verification and bilingual EN/FR consistency.

Each lands on Hugging Face under the base model's license when it's ready. Status updates: partnerships@cql.ca.

OPEN QUANTS

Have a model you want
quantized in the open?

We ship one or two community quants per cycle. We're looking for:

  • OPEN-WEIGHT BASEApache-2.0, MIT, Llama-compatible — nothing research-only.
  • REAL DEMANDAt least one team waiting to deploy it.
  • RECIPE-FITW4A16, NVFP4, or MXFP4 + vLLM target.
  • REGULATABLENo use-restrictions that conflict with regulated-industry deployment.

First-look review within two weeks. Selected requests get a public reproduction repo and a model card under canada-quant/.

Target quantization