One month, ten model lanes.
Every entry below was measured on a Mac mini M4 with 16 GB, in the order it happened. Newest first.
-
2026-09-17
DeepSeek V4.1 Flash lane live
A 466 GiB bundle prepared straight from the published safetensors shards, the first lane built without a GGUF source. Output verified. First light at 0.37 tok/s, entirely SSD-bound on the 16 GB mini; the levers to lift it are known and queued.
-
2026-09-12
Batched prefill, five steps
Flash Next's 578-token passage prefill went from 69.6 s to 36.7 s through five measured improvements to how the prompt is processed. Each step was verified before the next was built; one regression was measured and reverted the same hour.
-
2026-09-06
DeepSeek V4 Flash rebuilt, and the 27B as a mixture of experts
The lane rewritten from the model's published specification: 1.11 tok/s, output verified. Qwen 3.8 27B re-cut into a mixture of experts on its own weights, running half its experts per token at 3.07 tok/s with output unchanged. The old DeepSeek lane was deleted.
-
2026-09-05
One bundle format, the server bridge, and the old lane removed
One folder per model, one settings file per model, new bundles for gpt-oss and Gemma 4. The Studio's server gained streaming heads on the first byte, JSON errors before the stream, typed message parts, and Xcode's model bridge routing. Read-ahead ported to Flash Next: 3.18 to 3.95 tok/s. The original QPT bundle lane was deleted; everything is .qpx now.
-
2026-09-04
Companion ships. A lighter output stage.
Quadpoint AI Studio Companion became a native macOS app with the agent inside the bundle, the F5/F6/F7/F9 loop, and a verified two-round compile-error repair on a real project. The output stage of every token was made far lighter with the same answers, and Qwen 3.6 reached 16.17 tok/s, plus 7 percent.
-
2026-09-02
The kernel question closed; IO is the wall
Three GPU kernel variants, all verified, all within milliseconds of each other. The budget was elsewhere: 33 to 41 ms of every token was waiting on the SSD. Shipped read-ahead and the settings gear.
-
2026-08-31
GLM-5.3 Flash engine shipped, Downloader v1
A 163 GB bundle and a new engine lane. The internal Downloader gained Hugging Face and ModelScope search, a verified queue and a Library pane that reads each bundle's stamp.
-
2026-08-28
Stamped bundles
Every bundle became a self-contained file that records what it is and where it came from. Nine bundles across six model families were stamped in place. The Companion's patch lane went end to end: a local model wrote, broke and repaired a real SwiftUI app that built and ran.
-
2026-08-26
Qwen 3.8 Flash Next, download to full stack in one day
A 125B-parameter, 512-expert preview: census, a new slicer for raw i-quants, an 87 GB slice, and a full engine family with hyper-connections, n-gram hash memory, a sparse-attention indexer and gated delta attention. Verified at first light, 2.79 tok/s, then 3.18 after the decode IO war, then sparse attention to 8192 context the same night.
-
2026-08-24
DeepSeek V4 Flash first light
A 158 GB, roughly 290B-parameter frontier model prepared and running in the app at 0.13 tok/s in one evening, output verified. The next day's expert cache took it to 1.03 tok/s. Same night: Qwen 3.6 flash decode, 8.83 to 10.08 tok/s flat to 6k context.
-
2026-08-23
Qwen 3.6 Metal optimization, plus 93 percent
5.76 to 11.12 tok/s in one session on the resident 35B-A3B lane.
-
2026-08-20
gpt-oss 20B, spec to silicon in a day. Quality era.
Verified, about 19 tok/s, harmony-native with a Low/Medium/High reasoning selector, KV prefix reuse taking prefill from 5 s to 1.3 s. Sampled decoding with per-model official parameters and a min-p floor fixed the corrupt HTML that greedy decoding produced on quantized models. The play button that runs generated HTML in-app, and search-and-replace fix turns, shipped the same night.
-
2026-08-19
Multi-token prediction
Draft-and-verify through the bandwidth wall. The model drafts ahead of itself and the main pass verifies, so every step keeps the output unchanged.
-
2026-08-18
The .qpx format
Qwen 3.8 27B in a single mixed-precision file: 6.37 tok/s with verified output, 11.34 GB fully resident on the 16 GB machine.
-
2026-08-17
First inference
Slicer and engine validated on a 1.5B coder. Then Qwen 3 Coder 30B, coherent at 2.58 tok/s on the CPU alone, proving that routed-expert paging keeps about 1.9 GB of activated weights page-cached with no dense streaming. The Metal path reached 14.73 tok/s the same day.