ExLlamaV3 README feature list and JIT build note (read 2026-10-04)

Read 2026-10-04, used to fill the feature-comparison axes for [[exllamav3]] and to give that record its first build-cost figure. AXES CONFIRMED from the README's own bullet list: 'Generation - Continuous, dynamic batching, speculative decoding, multimodal support' - so continuous batching, speculative decoding and multimodal are all yes, stated by the vendor. QUANTIZATION: EXL3 'based on QTIP, plus 2-8 bit cache quantization', with AVX2/AVX512 CPU-offload support for large MoE on limited GPU. PARALLELISM: 'Flexible tensor-parallel and expert-parallel inference for consumer hardware setups' - expert parallelism is notable because it is rare among consumer-oriented engines. SERVERS: ExLlamaV3 ships no server of its own; the README names TabbyAPI as 'the official and recommended backend server' with an OpenAI-compatible API, embedding-model support and HF Jinja2 chat templates - so any OpenAI-compatible surface for this engine belongs to a DIFFERENT repository and multi-model / structured-output capability at the engine level cannot be answered from this README. BUILD COST, the one measured statement in the file: the PyPI package 'does not contain a prebuilt extension and requires the CUDA toolkit and build prerequisites', and both `uv sync` and `pip install .` 'build the package in an isolated environment where your torch is not visible, so they install the extension sources and compile them at first import (JIT, a few minutes once per torch version)'. Two facts a reader needs: the JIT compile is 'a few minutes once per torch version' - i.e. repeated per torch version, not per model - and the documented workaround is `pip install --no-build-isolation .` in an environment that already has torch, or a release wheel. The README also documents CUDA version selection as a build input: flavor extras cu124, cu126, cu128, cu129, cu130, cu132, minimum torch 2.6.0 and CUDA 12.4.