At least the Qwen3 30B A3B model can break an unknown fantasy word down from a set of tokens into single letters and count them correctly, as shown in the attached screenshot (running in my own WIP LLM inferencing engine, implemented in Object Pascal and x86-64 inline assembler).

Here a quick tinymembench benchmark performance comparison between the PasRISCV (Object-Pascal, FPC 3.3.1) and RVVM (C, GCC 14.2.1) RISC-V emulators.

Regarding how the instruction interpreter part in PasRISCV roughly works:

The beginning of network support in my Pascal 64-bit RISC-V PasRISCV emulator. 🙂 Although I still have a few routing issues here at the moment.

Good news! I've launched .masterknecht, a new free browser-based PWA tool for mastering audio using WebAssembly! Master your tracks with reference files. No ML, just efficient algorithms for fast mastering on your device. Try it & donate! masterknecht.klangknecht.com #audiomastering #PWA

PasVulkan now includes an external HTML5/JS-based tool for visualizing Vulkan memory allocations. Note that PasVulkan's custom Vulkan memory manager predates AMD's first public VMA release, addressing any questions about why VMA isn't used.