ARGMAX AI ARGMAX AI
RESEARCH LIBRARY

Flagship Technical Dispatches

Exhaustive first-principles architectural teardowns, LaTeX mathematical derivations, and verified runnable PyTorch/Triton implementations.

ALPHA: 96.4 arXiv:2603.14920 • 2026-10-19
24 min read

Dual-Chunk Recurrent Linear Attention: Sub-Quadratic Scaling at 128k Context

Softmax attention scales quadratically O(N^2) in sequence length, consuming over 580 GB of KV-cache memory across 32 concurrent 128k streams on 8x H100 clusters. Dual-chunk linear recurrence decomposes sequences into hardware-coalesced 128-token chunks, bounding memory cost to an invariant O(1) hidden state matrix during auto-regressive decoding while preserving full syntactic precision.

-89.7% KV VRAM · 3.2x Speedup | Linear Attention / State Space
Read Teardown & Code →
ALPHA: 95.8 arXiv:2502.12874 • 2026-10-19
22 min read

DeepSeek-V3 Multi-Head Latent Attention & FP8 Triton Kernel Deconstruction

Multi-Head Latent Attention (MLA) decouples key-value caching from head counts by compressing Key-Value projections into a low-rank latent vector (d_c = 512). Independent lab benchmarking confirms 73.2% VRAM reduction while matching standard Multi-Head Attention expressivity on Hopper H100 Tensor Memory Accelerators.

-73.2% KV VRAM · 2.4x Speedup | Latent Attention / MoE
Read Teardown & Code →
ALPHA: 91.8 arXiv:2410.02525 • 2026-10-19
20 min read

BitNet b1.58: 1-Bit LLM Architectures with Dynamic Precision Scaling

Every standard neural network computes matrix products via power-hungry floating-point MAC units. BitNet b1.58 constrains all linear layer weights to ternary values {-1, 0, +1}, turning matrix multiplications into integer additions and subtractions. Slashes dynamic inference energy by 82% and fits a 70B parameter model within 36 GB of unified memory.

-78.5% KV VRAM · 3.8x Speedup | Quantization / Integer Computing
Read Teardown & Code →