State of local AI, Oct 2026

In October 2026 an Apple Silicon Mac can run every model a small video editor needs without a server: a vision model to label clips, a model that calls tools, speech recognition with word timings, and one embedding space for text, frames, and audio. The best small options are task models measured in megabytes, from Cactus Compute and Desert Ant Labs, beside open general models from Google, OpenBMB, Liquid AI, and Qwen. Speech, sound, and image generation also run locally in seconds. Video generation is the exception: it works on a Mac, but a five-second clip still takes minutes.

This survey is written for Yapwing, a native Mac editor for captioned portrait videos that Hraness is building. Yapwing cuts a recording by editing its transcript, so it needs four models: vision for labeling clips, a tool-calling model, transcription, and multimodal embeddings. The four generation sections cover the media an editor might add. Each table lists the model, its size, license, the runtime that reaches a Mac, and what its authors report. Every figure comes from the linked model card or vendor page as of October 15, 2026; the vendors ran these benchmarks themselves, and numbers from different vendors are not directly comparable.

Two companies that build small task models

Cactus Compute ships one CPU engine and two models for it. Needle 3 is the tool-calling model: an 8 to 29 MB file that picks functions and fills their arguments, extracts typed fields from text, and returns text embeddings. Whistle is the speech-to-text model: 16.9 MB, with word timestamps and a speech embedding. Both load into the same Needle binary, which can transcribe a clip and answer it against your tools in one call. Both are Apache 2.0. Cactus also republishes Gemma 4, LFM2.5-VL, Qwen3.5, Parakeet, and Whisper in its own format.

Desert Ant Labs launched on September 8, 2026 with 18 single-task models for audio, vision, and text, 12 stable and six in beta, behind one Swift, Kotlin, and JavaScript SDK (launch post, Hraness reading digest). The team built them to replace cloud APIs in its video app, Detail, which makes the catalog unusually close to an editor’s needs: Voz for transcription, Align for word timestamps, Uhm for filler words, Clips for picking highlights from a transcript, Clear for audio cleanup, Who for speaker labels, and Eye for frame scoring. The models are source-available, not open source: the license is free below 100,000 monthly active devices per model and platform, requires a visible “Powered by Desert Ant Labs” credit, and forbids training a competing model on their outputs.

Vision for labeling clips

Pick: MiniCPM-V 4.7 1B for free-form clip descriptions. It takes video directly, it is Apache 2.0, and OpenBMB released it in October 2026 with GGUF builds for llama.cpp. When labels come from a fixed list, such as shot type, indoor or outdoor, or whether a face is visible, Liquid’s d1 models are a better fit: they return a calibrated answer to each question in one pass and never produce free text to parse. Gemma 4 E4B is the choice when a label depends on what is said as well as what is shown.

A model that calls tools

Pick: Needle 3 for turning a request into one of Yapwing’s editing commands. Its decoder is constrained by a grammar compiled from the tool schemas, so every call parses, and an off-topic request returns an empty list. Cactus also reports that fine-tuning on one product’s tools lifts each Needle depth by 18 to 36 points on DroidCall. Needle is weaker on broad benchmarks such as BFCL v4, so a task with several dependent steps needs a general model. Gemma 4 E4B is the smaller general option. LFM2.5-24B-A2B scores higher on Liquid’s tool benchmarks but has to hold 24B parameters in memory.

Transcription

Pick: Whistle, which Yapwing already uses. It is 28 times smaller than Voz, runs on the CPU with no other runtime, and returns the word timestamps that transcript cutting depends on. Qwen3-ASR 1.7B is the accuracy reference to test it against on real recordings. Two Desert Ant models fill gaps no recognizer covers: Uhm marks “uh,” “um,” and “hmm” to 20 ms without running speech recognition, and Align corrects the word timings from Apple’s SpeechAnalyzer, cutting average timing error by 65% on clean audio in Desert Ant’s tests.

Multimodal embeddings

Pick: EmbeddingGemma 2. The original EmbeddingGemma, released in 2025, embedded text only. EmbeddingGemma 2, released October 6, 2026, puts text, images, video, and audio in one vector space, so a typed query, a voice memo, and a clip can be compared directly (Hraness reading digest). Its vision and audio encoders load separately: text alone needs 270M parameters, and text with images needs 440M. Vectors truncate from 768 to 256 dimensions with little loss in Google’s tables. The alternatives either drop audio, like Qwen3-VL-Embedding, or forbid commercial use, like Jina’s omni models.

Speech generation

Pick: Pocket TTS when a voice has to match the speaker, because it clones from a short recording and runs on two CPU cores. Kokoro is the smaller choice for fixed narrator voices. The Open TTS Leaderboard scores intelligibility and speaker similarity with automatic metrics; it does not measure naturalness, so listen before choosing.

Music and sound effects

Pick: Stable Audio 3 Small-SFX for sound effects and Small-Music for short beds, since both run on a Mac CPU in a few seconds. Read the Stability AI Community License before shipping. ACE-Step 1.5 is the MIT-licensed option for full songs.

Image generation

Pick: FLUX.2 [klein] 4B for overlay images, because it generates and edits in one Apache 2.0 model and mflux runs it natively on Apple Silicon. Z-Image-Turbo is the alternative when realism matters more than editing.

Video generation

Pick: FastMetal-5B-QAD for generated b-roll. It is the only model here with published Mac timings at 720p, and it fits in 16 GB of unified memory. At 47 to 151 seconds a clip on an M4 Max, it suits a background job that imports the result as a new asset, not an interactive preview.

Limits

Every quality figure above is a vendor’s own measurement on public test sets. None of these models has been tested on Yapwing’s recordings. Desert Ant notes that clean audiobooks score 2 to 4% word error rate while meetings and podcasts score 10 to 13%, and that most real recordings sound like the second group. Licenses vary from Apache 2.0 to research-only, and several change terms at a revenue or usage threshold, so check the current license on the model page before shipping. Desert Ant Eye is in closed beta and has no published benchmark.