The front page carries only the newest release. These are the notes each version shipped with, kept here unchanged rather than overwritten by the one after it.
Qwen3.8-Flash-Next joins the same one-time purchase as DeepSeek V4.1 Flash rather than arriving as a second thing to buy. If you bought the purchase in 1.0.6, this model is already yours; there is nothing to buy again and nothing to restore. It runs with multi-token prediction, where the model drafts several tokens and then checks them in one pass.
The same split that made V4.1 possible. Most of this file is a per-layer embedding table — 95 GiB of it — read a row at a time and never held in memory, so what has to fit is roughly 45 GiB and not 147. On this M1 Ultra it is fully GPU-resident at 43.7 tok/s decode and 510 t/s prefill, measured on open prose with 71.7% of drafts accepted. Code runs faster, 47–51 tok/s; the slower figure is the published one.
A claim of 50–60 tok/s on a 64 GB Mac is why this was measured on one. It fits, at exactly 43.19 GiB planned. The claim is not reproduced: on a real M2 Max the best figure is 28 tok/s, on code, and plain decode is 16–23. That machine gives about 60% of the Ultra's decode and 45% of its prefill, which tracks memory bandwidth, and it accepts drafts at the same 71.7% — so the shortfall is the hardware, not the drafting. Published as a range, because this laptop lost 28% across a run.
Every kernel this architecture needs is now generated by our own compiler and verified against the reference engine — bit-identical logits across 2,200 positions, prefill, drafting, batched sessions and image embeddings — and it is ahead: +4.8–8.4% decode, +79% serving eight chats at once. Stated plainly: the app still runs this model on the reference engine, so none of that margin is in the rate above yet. It is the engine, ready, not a number you have.
The release that gave the app something to sell, and the release that had to learn the difference between what a model weighs on disk and what has to fit in memory before it could sell anything.
DeepSeek V4.1 Flash is unlocked by a one-time purchase in the app. Everything else in the catalogue stays free, and the weights themselves are MIT and free to fetch; what the purchase buys is the app's support for them — the verified download, the streaming setup, and an engine that can run the architecture at all. It unlocks on every Mac signed in to your Apple Account, with Restore beside the Unlock button.
Only about 152 GiB of that file is weights. The rest is a table the engine reads a row at a time from the file, never holding it in memory, so the model runs on a machine far smaller than its download. On this M1 Ultra, with the file on the internal SSD: 2.9 tok/s decode and 70 t/s prefill. The drive is part of the claim — the same file on an external USB SSD decodes at 0.54 tok/s.
A model whose table stays on disk was being judged by its download, which made a 366 GB file look impossible on any Mac and offered no way to run it. The model list now separates the two: what you download, and what has to fit. Without that distinction this release would have had nothing to sell.
Some models run only on the reference engine, which has no CPU-expert path and will not stream them from SSD. The app used to offer those modes anyway, and picking one meant a model that simply refused to load. Each row now offers only what its own engine can actually do, and says why the others are missing.
The release that gave the app eyes. A vision checkpoint, the engine work that let it ship without costing another model its speed, and a plain account of which hardware it has actually been run on.
DeepSeek V4 Flash Vision-Exp answers questions about an image you attach in the Playground. It arrives as one catalogue entry carrying two files — the model and a separate encoder that the model refuses to run without — 81.6 GiB together, resident on a 128 GB Mac. The attach control stays hidden unless the model being served can actually see, so the app never offers a picture the engine would only refuse.
A new attention kernel reads its tiles straight from device memory instead of staging them. Encoding the same image takes 0.398 s against 1.243 s — 3.12×, measured in the engine the app actually ships rather than in a patch alongside it. On by default.
Vision needed a newer engine, and that engine served GLM-5.3-Flash too — where it cost 44.6% of prefill on prompts under 4096 tokens. A 2048-token prompt would have gone from about 10.7 s to 19.3 s before the first word. One line restores the path GLM had: 194.5 t/s, all of it back, with vision untouched. Both models are served by a single engine.
Vision is measured in full on Metal, and verified on ROCm on real gfx1151 hardware — the two give word-for-word identical answers. CUDA has not been run at all, not even compiled, because the only toolchain within reach rejects the architecture it needs. Said here rather than discovered later.
GLM-5.3-Flash — 320B parameters, 18B active per token, 96.5 GB. The first GLM this app can hold entirely in memory on a 128 GB Mac rather than streaming it from disk: 24 tokens/sec measured on an M1 Ultra. GLM-5.2 is retired — at 217 GB it never fit on a machine that size and only ever streamed. A copy already on your disk still runs.
Qwen3.8-Flash-Next — 125B parameters with 6B active, 78.9 GB, hybrid linear and sparse attention. 31 tokens/sec on the same machine, which makes it the quickest of the large models in the catalogue, and the smallest of the three that need a 128 GB Mac.
Qwen3.8-Flash-Next was listed at roughly an eighth of its real speed, its memory-saving modes were shown as unavailable when they are not, and the bundled engine was too old to load it at all. All three are fixed — which is why it appears in this release with a number worth reading.
Models served by llama.cpp now use their own prompt format instead of a substitute, so they behave the way their publisher intended; and a temperature of 0 returns the same answer every time rather than a sampled one. The licence credits now name every model the app can download.
A Generate Video screen runs MiniMax H3 on your own GPU: seconds of video, with synchronised audio, from a sentence — or from two anchor frames, where you give the first and the last and the model fills the middle. Same rule as the text models: the weights are yours, the frames never leave the Machine.
The engine still binds loopback by default. You can now hand it to your phone or a colleague — but only by turning LAN serving on this session, and only with an API key you mint in the app. The key is shown once; the Mac keeps its SHA-256 and nothing else. The engine refuses to open up without one.
Qwen3.8-27B, now the recommended Qwen — dense, Apache-2.0, 256K context. Muse Glimmer 30B from Meta, 30B dense in 17 GB, whose hybrid attention keeps a full 128K session near that size too. MiniMax H3 for video. And DeepSeek-V4-Pro-0813 — 1.6T parameters, 850 GB, listed and honestly priced rather than hidden.
DeepSeek Harness (dsh) and Grok Build join the
one-click wiring. The rankings row became data too — add the leaderboard you actually
trust, alongside the ones we shipped.
1.0.2 shipped before the front page carried a release band, so there is no poster copy to keep. What it shipped with is the App Store text, reproduced as written.