MLC-LLM

'mlc-compiled' is MLC's own artifact format; q4f16/q4f32 are the published quantization variants (group-quantized weights with f16 or f32 scales). WebLLM (mlc-ai/web-llm) is the browser deployment of the same compiler and is recorded here as part of MLC-LLM rather than as a separate engine record. ||| MAINTENANCE (dated fact, GitHub API, 2026-10-04): last commit 2026-10-04 (0 months), NO RELEASE OBJECT EXISTS in the GitHub API (the tags endpoint returns a list but no releases/latest, so release publication dates are unverified for this project), 164 contributors, open_issues_count 349 (issues PLUS pull requests), NOT archived, 23205 stars - a high star count with no release history, which is a distinctive maintenance shape. DEPLOYMENT: NO EVIDENCE OF PRODUCTION DEPLOYMENT FOUND as of 2026-10-04 - see [[ev-production-evidence-is-scarce-and-nearly-always-single-site]]. Note that this is a weaker negative than for a small library: the README publishes an offline batch-inference throughput TUNING GUIDE with a worked example (a log line showing 'Decode batch. #running-req: 233, #token: 370959, token usage: 0.82, cuda graph: True, gen throughput (token/s): 4594.01, #queue-req: 317'), which is evidence that operators tune this engine seriously but is explicitly a configuration guide, not a report of anyone's deployment. FEATURE AXES - THIS IS A RECORDED NEGATIVE AND IT IS A NEGATIVE ABOUT THE SOURCE, NOT THE ENGINE: a case-insensitive scan of the README for speculat|eagle|mtp|grammar|json.schema|structured|chunked|prefix cach|radix|disagg|multi.model|multimodal|vision|audio returned ZERO matches. Speculative decoding, chunked prefill, disaggregated P/D and structured output are therefore UNSTATED-AND-UNVERIFIED from this source, and prefix caching - which the record asserts - is unconfirmed rather than contradicted. The README is a short project overview; the authoritative per-model surface is the model library, which was NOT enumerated, so absence of evidence here is not evidence of absence and none of these axes should be recorded as NO on the strength of this. ONE DISTINCTION TO PRESERVE: the record's existing claim of one engine behind Python/REST/JavaScript/iOS/Android is about serving SURFACES, not about serving several models at once, so it is not an answer on the multi-model axis. BUILD COST: no published figure found for MLC LLM's model compilation, which is a per-model artifact build by design and would be the cost to record; this record does not estimate it.