Patent lawyer shares data-engineering tips for classifying 3.5M US patents with Nemotron 9B on RTX 5090
A patent lawyer who started coding in Dec 2025 posted a follow-up on classifying 3.5M US patents using Nemotron 9B on a single RTX 5090. Key data-engineering tips include that PatentsView moved to the USPTO Open Data Portal in March 2026, old S3 download links are dead, and abstracts are now in a separate table. He also advises running ANALYZE after bulk-loading to avoid query plan issues.
Entities
Related
Developer compresses GLM-5.2 MoE to run on single RTX 3090 via 79 experiments
A developer conducted 79 experiments to compress GLM-5.2, a 337 GB MoE model with 75 sparse layers and 256 routed experts, to fit on a 24 GB RTX 3090. The approach uses per-expert codecs, a batch pipeline over all MoE layers, and a patched llama.cpp runtime that loads codec-native expert binaries at inference time. The MIT-licensed repository documents the method and findings on expert similarity.
Rust+CUDA LLM inference engine for RTX 5090, optimized with NVFP4, MoE, and speculative decoding
A from-scratch LLM inference engine written in Rust and CUDA, specifically tuned for the RTX 5090 Laptop GPU (Blackwell sm_120a). It supports NVFP4 quantization, mixture-of-experts (MoE), and multi-token prediction (MTP) speculative decoding, achieving up to 1.6x speedup over llama.cpp on target models. Designed for developers who need maximum inference performance on a single high-end laptop GPU.
Developer adds Nemotron Puzzle 75B support to mlx-lm, benchmarks 4-bit vs 5-bit quantization on M2 Max
A developer contributed native `nemotron_h_puzzle` support to mlx-lm (PR #1535) and ran benchmarks on a 64GB M2 Max. Comparing 4-bit vs 5-bit expert quantization (both with 6-bit dense layers, BF16 output head, group size 64), the 4-bit variant used 42.03 GiB checkpoint (49.68 GB peak memory) and achieved 14.27 tok/s, scoring 24/30 on local task checks, while 5-bit used 49.88 GiB (58.12 GB peak) at 10.53 tok/s scoring 21/30.
LLM hardware recipe database with filters and community usage tracking
A community-driven database that lists which LLM models run on which hardware, with performance details. Users can filter by hardware, submit new recipes, and mark which recipes they actively use to show popularity. It solves the problem of finding compatible model-hardware combinations for LLM deployment.
User tests GLM 5.2 on consumer hardware, finds performance comparable to Claude and GPT
A user tested GLM 5.2 on a standard computer and was impressed by its capabilities and security, finding them similar to Claude or GPT. They began converting the model to int4 and exploring MTP usage to avoid out-of-memory errors.