Release writeups and engineering notes from the lab — recipes, evaluation methodology, and the walls we hit along the way. The same rigor that goes into the model cards, in long form.
Our Canadian legal-reasoning benchmark now has a live leaderboard at canlegalbench.com: 14,547 primary-source records, 17 task types, 14 jurisdictions, 31% in French, five frontier models scored closed-book. Claude Fable 5 leads at 80.1, with GPT-5.6 Sol and Gemini 3.1 Pro close behind. No model clears 60% on citation resolution, citation holdings, bijural equivalence or issue spotting.
DFlash2-G, our third self-trained drafter, is the first to beat the widely-used reference drafter on acceptance, outside the noise band. On 8× H100 our W4A16 quant beats NVIDIA's own NVFP4 checkpoint on the same rig (+5.1% at c256; +47.8% single-stream against NVFP4 run without speculative decoding) and posts the model's first formal vision scores: MMMU 0.747, OCRBench 888. On 2× DGX Spark, NVIDIA's NVFP4 stack is ahead, and we say so.
We trained a DFlash2 block-diffusion speculative-decoding drafter for GLM-5.3-Flash from scratch — 350,260 self-generated samples, no third-party drafter weights or traces, Apache-2.0. It beats our previous best in every same-rig A/B cell on 2× DGX Spark, reproduces on H200 within 0.3%, and is now our serving drafter-of-record — with the 2× Spark serving stack open-sourced alongside.
We quantized Z.ai's multimodal GLM-5.3-Flash to 4-bit weights (177.7 GiB, −70%) while keeping its BF16 MTP draft head and vision tower. It serves the full 1M-token context on 2× DGX Spark and scores 85.0% on AIME 2026. Updated 30 Sep with corrected NVFP4 comparisons.
We quantized Tencent's Hy3 (295B MoE) to 4-bit weights while keeping its MTP draft layer in BF16. It matches the FP8 release on quality at 57% of the footprint, runs +40% faster single-stream than the same-scheme quant that shipped without MTP, and takes the high-concurrency throughput crown among 4-bit Hy3 quants.
We quantized GLM-5.2 (744B MoE) to 4-bit weights while keeping its MTP draft head in BF16. It matches the FP8 release on quality, fits on four H200s instead of eight, and is the fastest popular 4-bit GLM-5.2 quant in the interactive serving regime.