Open Source Reports
August 2026

The month went into two new pure-Swift libraries, both ports of a reference implementation in another language. swift-pii finds and removes personal data — a behavioural port of Microsoft’s Python Presidio, down to the spaCy pipeline it depends on, including the dependency parser. swift-anydoc converts Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV and PDF to Markdown — a port of Rust firecrawl/anydoc, with the Rust binary used as an oracle so that “correct” means byte-identical rather than “the tests pass”. Neither has a tagged release yet; both are public and working.

LMPlayground shipped 1.9.1 — three new reasoning models that can also use tools, and seven fixes, four of which were found by a new macOS harness that runs the whole model catalog against real weights instead of waiting for a phone to prove something is broken.

Both Spark projects sat still in August; their 1.3.1 landed in early September and belongs in the next report.

LMPlayground

• 1.9.1 — Added LFM2.5 2.6B, SmolLM3 3B and MiniCPM5 1B — reasoning models that can also search the web, fetch a page and run JavaScript. LFM2.5 1.2B now ships the Instruct build (at that size, thinking before every reply costs more latency than it returns), and Qwen 3.5 moves from Q3_K_M to IQ4_XS — about half the quality loss at the same speed for 3% more download. Fixes: SmolLM3’s tool calls arrived as raw XML because its published chat template never rendered them; image turns passed an uninitialised length to the tokenizer (undefined behaviour on every image); Gemma 4 images are ~4.5× faster on a Q8_0 projector instead of the BF16 one it shipped with, which has no optimised ARM kernel; long chats could come back empty when a stale KV cache skipped context compaction. The engine moved into a new :llamacpp JVM module so it runs without a device, which is what the new :model-harness uses to sweep the catalog on macOS — templates, tool round trips, thinking, vision, history replay, context overflow — and where four of those fixes came from. llama.cpp at b10327.

swift-pii

• New project — Presidio-compatible PII detection and anonymization in pure Swift, macOS and Linux, no Apple-only frameworks and no Foundation in the core. 88 recognizers: 17 on by default (cards, IBANs, e-mail, phone, SSN, passports, dates) and 71 country-specific ones opt-in. Detection is scored rather than binary — a Luhn-valid card is 1.0, a bare phone shape is 0.4, a supporting word nearby raises it. Anonymizer then redacts, masks, replaces, hashes, or encrypts reversibly. Names, places and organisations need the language model, and spaCy’s en_core_web_sm is bundled (12 MB) so that path needs no download. That meant porting the pipeline itself: tokenizer, tagger, the edit-tree lemmatizer, NER, and the dependency parser — the last one only because spaCy forbids an entity from spanning a sentence boundary and takes those boundaries from the parse, which is what closed the final NER gaps. Measured against spaCy: 50,⁷³¹⁄₅₀,731 dependency heads and labels, 2,⁴¹⁴⁄₂,414 sentence boundaries, exact tokenization in all seven non-English languages, and 28,669 recorded parity cases overall. Upstream is pinned to an exact commit and CI refuses artifacts stamped with any other. Also ships a CLI whose exit status is meaningful — 1 when PII is found — which upstream’s cannot do.

swift-anydoc

• New project — Documents to GitHub-Flavored Markdown in pure Swift: Word, PowerPoint, Excel (including the legacy .doc/.ppt/.xls record formats), OpenDocument, RTF, EPUB, CSV and PDF, across 20 extensions. Package.swift fetches nothing; zlib is the single system link, and import Foundation is banned inside the library and checked in CI, which builds on Linux where those frameworks do not exist. ZIP, MS-CFB, a namespace-aware XML pull parser, legacy codepage tables and the whole PDF stack are written in-repo. The Rust reference is treated as an executable specification and built locally as an oracle: 57 of 58 reference fixtures convert byte-identically, plus 168 generated adversarial PDFs byte-identical against the reference binary, 1,384 corruption mutants with no crashes or hangs, 265 hand-built adversarial documents, and 1,793 tests. PDF conversion runs at about 1.5× the Rust reference. .xlsb is deliberately unsupported — the reference reaches it through a third-party crate and no fixture exercises it on either side, so a port would have nothing to check it against.