The biggest open weights ever landed today; the best app ships none.

A macOS menu-bar dictation app: press a shortcut, talk, press it again, and the text lands in whatever field you were already in. What makes it worth a slot on a day about open weights is that it carries no model. macOS 26 added two APIs, SpeechAnalyzer and SpeechTranscriber, that do on-device streaming speech-to-text against models the OS ships and manages, so Yap is roughly three thousand lines of native Swift in a 4 MB app that loads nothing into memory before the first word and idles around 60 MB. No API key, no per-minute cost, no browser engine, no 600 MB of Whisper weights parked in RAM. The README cites an independent benchmark across 5,559 LibriSpeech clips putting Apple's model at 2.12% word error rate on clean audio and 4.56% on noisy, against Whisper Small's 3.74% and 7.95%, at roughly three times the speed.

The thread surfaced the fact the README undersells. The obvious objection is that macOS already has dictation built in — and it does, except Apple's own disclosure says it sends voice input, contacts, and location to Apple's servers. Yap calls the on-device API directly and makes zero network calls, and it is open source, which is the only reason you can confirm that rather than take it on faith. Invert that and you have the shape of this whole edition: the weights are a black box you were given for free, and the thing being opened is the app wrapped around them. Built by Frigade, whose auto-generated-MCP tool was a pick here on 07-10, and the 07-20 voice-stack watch is now resolved — though not in the form that watch predicted. It guessed a coding-agent harness would wire push-to-talk natively; this is OS-wide dictation you point at any harness. Third time this month a watch has landed above the layer it named.

Reach for it when you dictate prompts to coding agents all day and resent paying a subscription for it. Delete the paid dictation app and the gigabytes of local speech weights it needed to justify itself. Tradeoff: macOS 26 or newer only, and Apple's model is a box you cannot swap, inspect, or fine-tune — one commenter correctly noted that "works with wifi off" proves a local path exists, not that the network path is never used; NVIDIA's Parakeet TDT v2 still benchmarks ahead on accuracy; and there is a name collision with an unrelated finnvoor/yap that wrapped the same APIs last year. MIT, 180 stars since 07-23, signed and notarized .dmg shipped this morning as v0.1.2. The Homebrew tap in the README was broken at launch and fixed within the hour.

Background removal, released as two things rather than one: a trained model and the library that trains it. FeyNoBg extends the BiRefNet architecture by widening the third feature-extraction stage from 18 blocks to 24 — a modest expansion chosen specifically so the pretrained weights survive it — and the authors claim the best published S-measure on four of eight benchmarks, leading on the high-resolution and ultra-high-resolution sets at 0.983 and 0.981. It is a pip install nobg away, or uv add nobg, with the checkpoint on Hugging Face under MIT.

The half that matters more is the library. Alongside inference, batching, and half-precision GPU paths, nobg ships fine-tuning on custom data, checkpoint re-parameterization, and push-to-hub — which is the $500-fine-tune argument from today's news pool with the tooling attached. Whoever wrote that catalog-review post had to assemble their own training loop; this is the same move for a narrower task, packaged. It also quietly answers a question the open-weights debate never asks: a 263M-parameter specialist beating general models at one job is not a story about frontier capability, it is a story about the size of the job. Most jobs are this size.

Reach for it when you are paying a per-image API to cut out product photos, or when a generic cutout model keeps failing on your specific kind of image and nobody will retrain it for you. Delete the background-removal SaaS line item. Tradeoff: the model was trained on converted binary masks rather than soft-opacity supervision from matting datasets — their own disclosure, and precisely where cutouts fall apart, on hair, glass, and motion blur; "best on four of eight benchmarks" also means it loses on the other four; and the repo is three weeks old at 70 stars, so the fine-tuning path is documented rather than battle-tested. Library Apache-2.0, weights MIT.

Drop in a video, get a transcript with per-word timestamps and speaker labels, delete words from the text, and the corresponding footage is cut. Export a frame-accurate MP4 without the file leaving your machine. The stack is entirely client-side and worth reading as a status report on how far that has come: transformers.js running whisper-base_timestamped or whisper-small_timestamped on WebGPU with a WASM fallback, inside a Web Worker; pyannote-segmentation-3.0 in ONNX for the speaker labels; multi-threaded ffmpeg.wasm for extraction and export. 231 stars in two days, and a hosted demo on GitHub Pages that needs no account.

Two corrections to its own headline, both of which make it a better example rather than a worse one. The Show HN says a weekend; the commit history says about four hours with Cursor, and the author confirmed it in the thread. And "fully offline" means offline after the first run — the models come down from the Hugging Face Hub the first time you transcribe, roughly 200 MB for Base or 600 MB for Small, then cache in browser storage, and the app fires anonymous Google Analytics that fails silently when disconnected. The author states both in the README, which is more than most. What he does not state is that there is no LICENSE file in the repository, so the "open-source" in the title describes code you can read, not code you have been granted rights to use. That is the second edition running with a repo trending as open source that ships no license, and at four hours of build time the omission is less hypocrisy than an artifact of the timescale.

Reach for it when you edit interviews or podcasts by deleting sentences and the subscription for that privilege has started to annoy you. Delete Descript at $24 a month. Tradeoff: the missing license is a real one until somebody adds a file; the thread's economic objection is fair on its own terms — you do not recoup four hours of senior engineering time on a $24 subscription, and the actual return is being able to add the feature the vendor will not; there is a name collision with ReScript, the OCaml-derived language; and browser-tab Whisper on a long recording is a different experience on an M-series laptop than on the machine in the conference room.

The kicker, because the ladder has to end somewhere honest. TokenSpeed is an inference engine built for agentic workloads, aiming at TensorRT-LLM performance with vLLM usability, and its headline achievement this month is Day 0 enablement for Kimi K3 — the model card's own deployment section points at it. The engineering is not vibes: a local-SPMD modeling layer whose static compiler generates collective communication from module-boundary placement annotations, so nobody hand-writes parallelism; a split control plane in C++ and execution plane in Python, with request lifecycle and KV cache ownership encoded as a finite state machine whose safe cache reuse is enforced by the type system at compile time; and a pluggable kernel registry including one of the faster MLA implementations on Blackwell. There are PyTorch engineering blog posts behind the kernel work and a claimed 580 TPS on Qwen3.5-397B-A17B.

This is the pick that makes the edition's argument instead of decorating it. K3's weights are free, ungated, and downloadable by anyone; the software required to serve them at the numbers everyone is quoting is also free, MIT, and on GitHub. And the overwhelming majority of people reading this still cannot run the model, because the Pareto curves in that README are measured on B200s. Open weights are a license, not access. Telnyx's $2.70 and $13.50 per million is not rent extracted on a closed model — it is the market price of a machine, charged for an artifact that costs nothing.

Reach for it when you are actually serving open-weight models at production concurrency and the gap between vLLM's ergonomics and TensorRT-LLM's throughput is costing you real money. Delete the metered API in front of weights you already have. Tradeoff: every performance curve is the project's own, measured against TensorRT-LLM on hardware most teams do not have, and the honest reading of Day 0 K3 support is that a 2.8T model at 1M context is a datacenter problem no engine turns into a laptop problem. LightSeek Foundation is also young — the org dates to October 2025 with seven public repos — and TokenSpeed itself cut v0.1.0 four days ago, which is an early version number for something now sitting on the critical path of a flagship release.

Also worth knowing. alibaba/skill-up (Apache-2.0, 226 stars, Go, v0.7.0) is the artifact the 07-17 vendor-skills watch asked for by name — the eval, published first-party. Declarative YAML cases run across Claude Code, Codex, Qoder, and qwen_code, judged by rule, script, or agent, emitting Anthropic-compatible grading.json and benchmark.json plus JUnit XML for CI; the skill-upper skill it ships reads failed reports and repairs the eval suite through conversation. It installs with npx skills add. Same day, HKUDS/OpenSpace (MIT, 7,127 stars) trended on the same verb — retrieve, evaluate, evolve, across Claude Code, Codex, OpenClaw, Hermès, and nanobot — so the watch resolved in two forms at once, the same way the doorbell watch did on 07-23. Let's Seal (Apache-2.0, 134 stars since 07-16, 87 points) wants to be Let's Encrypt for document proof: one SEAL standard, signatures native to each format so any standard validator checks them (PAdES inside PDFs, C2PA in media, detached CAdES elsewhere, in-toto and DSSE for build artifacts and SBOMs), timestamps anchored to Bitcoin via OpenTimestamps, an RFC 6962 transparency log, free forever, and self-hostable under your own CA — with an unusually straight boundary statement that it proves integrity, time, and issuing certificate, and does not assert identity. Off today's axis; the artifact-signing leg is the one to watch. microsoft/flint-chart (MIT, 2,369 stars, +218 today, v0.4.0) is our 07-10 pick re-trending, disclosed. MCP-Billing (79 votes, €79 one-time) is the 07-02 agent-billing watch landing as a self-hosted Next.js boilerplate rather than a protocol — OAuth 2.1 with PKCE, key rotation, usage-based Stripe, Redis rate limiting — with the metering core given away as MIT on npm as mcp-metering. And agentscope-ai/QwenPaw took the day's biggest star move at +818 to 29,492, noted for magnitude only: it is an established personal-assistant harness from February, not a launch.