Mistral 7B

A 7B model competitive with far larger models is wanted, but inference cost at that size is set by attention cost over long contexts and by KV-cache footprint.

Two inference-motivated architectural choices, both stated as such in the abstract: GROUPED-QUERY ATTENTION (GQA) explicitly 'for faster inference' - it shares key/value heads across query heads, shrinking the KV cache and the per-token KV bandwidth; and SLIDING WINDOW ATTENTION (SWA) 'to effectively handle sequences of arbitrary length with a REDUCED INFERENCE COST' - bounding the attended window caps the KV cache independently of sequence length.

No speedup multiplier. Capability claims: Mistral 7B v0.1 outperforms LLaMA 2 13B across all evaluated benchmarks, and LLaMA 1 34B in reasoning, mathematics, and code generation. Released under Apache 2.0, which is itself the adoption fact.

ADOPTION: in-production and about as widely deployed as any open model - Apache 2.0 weights, the default local model for llama.cpp/vLLM/LMStudio, and the base of a large derivative ecosystem. VENUE NOTE: no academic venue; this is a model release report, and the arXiv comment points to the Mistral news post rather than a venue, so venue = arxiv-preprint with year = ArXiv v1 (Oct 2023). INCLUDED IN AN INFERENCE SLICE DELIBERATELY, because GQA + SWA are the two most widely SHIPPED inference-cost mechanisms of the 2023-2024 generation and they are more consequential in aggregate than most of the speculative-decoding papers here - SWA is the reason a 7B model can serve long contexts at bounded memory, and GQA is the reason its KV cache is small enough to batch. Anyone attributing decode cost to architecture should start here rather than with a speculative-decoding paper. DANGLING REF STILL OPEN (2026-10-04 paper-read pass; supersedes the earlier 'LEFT UNRESOLVED' note, which reached the same conclusion but only from this record's own text - this time the PAPER itself was read). No A100 and no accelerator named anywhere in the paper. It is a model report whose hardware content is a deployment note ("using the vLLM inference server and SkyPilot", "We are grateful to CoreWeave for their 24/7 help in marshalling our cluster", "the operators of Leonardo"). Source: https://arxiv.org/pdf/2310.06825 . Left dangling; the paper does not state what it trained or benchmarked on. AFFILIATIONS FILLED 2026-10-04 (paper-verification pass): from the Semantic Scholar graph API (paper/arXiv:2310.06825), which is the only machine-readable source that carries author affiliations — the arXiv Atom feed returns author names only, and no ORCID is exposed by either API. One known upstream typo corrected: "Tel Aviv Univesity" -> "Tel Aviv University". S2 affiliation coverage is partial (120 of 249 corpus records carry any), so these lists may be incomplete rather than exhaustive. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'unknown'. `venue` is 'arxiv-preprint', which means no peer-reviewed venue has been established for this work as of the verification date, so the track genuinely is not known. 'unknown' is used here rather than leaving the field null or guessing 'main': the schema separates 'unknown' from 'main' precisely so that an unprereviewed preprint is not silently counted as main-track work. This annotation is what lets a downstream filter distinguish 'peer reviewed, main track' from 'not peer reviewed' without inferring it from the venue string.