Quintus is a compact 1.7B assistant model that improves over Qwen3-1.7B-Instruct on key reasoning and commonsense benchmarks, measured by +4.5pp on GSM8K flexible, +5.4pp on ARC-Challenge acc_norm, +5.4pp on WinoGrande, +6.6pp on MBPP, and +3.5pp on PIQA acc_norm. The model was built by distilling Qwen3-8B into Qwen3-1.7B through a two-stage pipeline using full-vocabulary online KL divergence, sequence packing, token-chunked KD loss, targeted SFT, strict evaluation controls, and public weight-audit tooling.

Keiro retrofits a Sparse Mixture-of-Experts architecture into Qwen2.5-3B. A Top-2 dynamic router activates 2 of 8 LoRA experts per transformer block, expanding effective capacity while keeping active compute identical to the dense baseline. The residual design — frozen FFN + routed Rank-16 LoRA adapters — adds only 19.46M trainable parameters (0.63% of total) while retaining 95.4% of the base model's GSM8K mathematical reasoning capability. Engineering challenges included resolving a CUDA race condition in index_add_ with duplicate Top-K indices, a BFloat16 cumsum upcast mismatch in the coalesce path, and a 4.7× autoregressive inference bottleneck diagnosed and addressed by bypassing capacity buffers for single-token generation. Benchmarked via EleutherAI lm-evaluation-harness (HellaSwag −0.13%, ARC-Challenge −0.17%, GSM8K −3.19%).

Speculative decoding for LLM inference — a small Qwen 1.7B draft model races ahead, predicting future tokens that a larger Qwen 8B target then verifies in a single parallel pass. The result: 1.30× faster generation on an A100 at ~56.5% acceptance rate across mixed-domain prompts, with a rejection sampling pipeline that mathematically guarantees the output distribution stays identical to the target model.

FlashTile is a reference PyTorch implementation of Flash Attention (V1/V2) and KV-cache-efficient inference variants (GQA/MQA). It uses block-wise tiling, online softmax, and recomputation-based backward passes to reduce attention-score storage from O(N²) to O(N). The project includes benchmark and validation artifacts from A100 testing, an archived H100 cross-check, and a forward-only Triton kernel for performance comparison.

Substrata9 — a lightweight, pure-Bash toolkit for deep Linux process introspection. Built entirely without compilation or external dependencies, it mines raw /proc filesystem data to surface memory maps, file descriptors, process hierarchies, and runtime anomalies. Its modular CLI utilities emit JSON output for automation, slotly cleanly into observability, debugging, and forensics workflows.