Skip to content

Releases

Astronomical 0.2.123

Choose a tag to compare

@aosama aosama released this 21 Nov 01:21
1840d7b

What's new

  • Astronomical now ships with a skill for AI agents. Coding tools and chat harnesses connected to Astronomical are told where it is, so their agents can learn on their own how to draw images, read text aloud, describe pictures, run decision models, connect tools and troubleshoot Astronomical. Nothing needs installing.
  • The skill keeps itself current. About once a day Astronomical checks the website for a newer skill, verifies every file against its published SHA-256 digest, and swaps it in whole. Nothing about your Mac or its use is sent.
  • Two new settings in Observatory, under "Astronomical skill": one tells models about the skill, and one keeps it up to date from the website. Both are on by default and apply at once without unloading the model.
  • astronomical speak now speaks aloud on your Mac's speakers. Pass --output FILE.wav to save the audio instead.

Fixes

  • The menu bar no longer shows "Not running" while a speech or decision model is loaded.

Compatibility

  • astronomical speak without --output no longer saves a file. Scripts that relied on the saved file should pass --output.
  • Your settings, conversations and downloaded models are kept. The new skill settings section is optional.
  • Requires macOS 14 or later on Apple silicon.

The app and disk image are signed with Developer ID, notarized and stapled by Apple, and the update feed is signed.

Astronomical 0.2.122

Choose a tag to compare

@aosama aosama released this 20 Nov 11:51
196f002

What's new

  • A quiet menu bar. The Astronomical icon now stands alone by default. While a prompt is read or a picture is drawn, the icon itself shows how far along it is. If you'd like the old progress and speed text beside it, switch it on in Observatory's Settings or from the ⋯ button in the menu.
  • A simpler menu. The menu that opens from the icon now shows your model, what it's doing, how fast it reads and writes, how much memory it uses, and your memory limit. The detailed memory breakdown and build information are still there: switch on "Show technical details in the menu".
  • A memory slider. Set the most memory a model may use by dragging. A mark shows the loaded model's size, and a sentence tells you what your choice leaves for macOS and your other apps, or warns when part of the model will be read from the SSD. Drag it all the way right to use whatever your Mac allows.
  • Reading and writing speed. The menu shows how fast the model reads your prompts and writes its replies, averaged over the session, with the live speed while a reply is under way.
  • A clear sign when something's wrong. If Astronomical's engine stops answering, the icon shows an exclamation mark and "Not running", and the menu offers Restart.
  • Slightly faster prompt reading on mixture-of-experts models held fully in memory, with identical replies.

Fixes

  • The session's reading and writing speeds now weigh each request by how much it did, so a very short reply no longer skews them.

Good to know

  • Your conversations, models, settings and saved prompt work are all kept.
  • Requires macOS 14 or later on Apple silicon.
  • The app is signed with the project's Developer ID and notarized by Apple.

Astronomical 0.2.121

Choose a tag to compare

@aosama aosama released this 20 Nov 02:21
8fcc3d6

Enhancements

  • The 6-bit vision Ornith 1.5 35B-A3B now drafts out of the box. The download catalog adds Ornith-1.5-35B-A3B-MLX-6bit-vision-DFLASH2: the 6-bit vision build with the DFlash2 draft model packed in, every file byte-identical to its source. It takes about 28 GB on disk and about 29.5 GB of model memory, so a 48 GB M5 Mac holds the model and its drafter whole. Measured against the plain 6-bit vision build on such a Mac at temperature 1.0, short replies decode about 1.16 times faster (89.4 against 77.3 tokens a second); after a 10,011-token prompt the two run about even. Macs without an M5 GPU, and memory limits too low for the drafter, decode exactly like the plain build. The plain 6-bit vision build stays in the catalog.

Compatibility

  • Existing configurations, installed models, conversations and settings are preserved; the configuration schema is unchanged.
  • macOS 14 or later on Apple silicon.

Distribution

Signed with a Developer ID under Hardened Runtime, notarized by Apple and stapled; the Sparkle update is EdDSA-signed over the final bytes and served from a signed feed.

Tracked in #2383.

Astronomical 0.2.120

Choose a tag to compare

@aosama aosama released this 20 Nov 01:00
5470c0f

What's new

  • Qwen 3.8 27B now answers about twice as fast on M5 Macs, straight from the catalog. The download catalog offers it as Qwen3.8-27B-MLX-4bit-vision-DFLASH2: the 4-bit model with its vision intact and a small draft model built in. The draft model guesses the next few words and the large model checks them all in one step. Every word is still the large model's own choice. After a 10,000-token prompt it writes 35 to 37 tokens a second, where the plain 4-bit build wrote 16.7. On Macs without an M5 GPU, with less memory, or for a message with a picture in it, it runs just like the plain build.
  • The catalog now offers that build as its only 4-bit Qwen 3.8 27B. The two 2-bit ternary 27B builds remain for Macs with 8 GB of memory.
  • Every page of the website shares one header and footer.

Under the hood

  • The model-serving code is split into smaller, focused parts, with no change in behaviour.
  • Release and website checks moved into the project's own typed tooling.

Compatibility

  • Your models, conversations and settings are kept. A model you downloaded earlier from a retired catalog entry stays installed and usable.
  • Requires macOS 14 or later on Apple silicon.

Astronomical is signed with a Developer ID, notarized by Apple, and delivered through a signed Sparkle update feed.

Astronomical 0.2.119

Choose a tag to compare

@aosama aosama released this 19 Nov 06:29
b4e465a

Fewer storage reads while streaming

When a mixture-of-experts model is bigger than the memory you give it, Astronomical reads the experts each word needs from the SSD. This release keeps whole only as many layers as still leave every other layer a fair share of memory, and sizes those shares from the memory they actually have, so far fewer experts have to be read again.

  • Ornith 1.5 35B-A3B 8-bit Astro Atlas at 17.6 GB: 21 to 49 experts read from the SSD per word, down from 71 to 109, and 32.1 tokens a second with nothing held in the macOS file cache, up from 22.2.
  • The 4-bit OptiQ build at an 11 GB limit: 33.3 tokens a second, up from 25.1.
  • Replies are word for word the same as before.
  • Each reply's hits, misses and bytes read from the SSD now show in /v1/experts/stats and in the log.

New in the catalog

  • Ornith-1.5-35B-A3B-MLX-8bit-vision with its DFlash2 drafter, for draft trees on the highest-precision Ornith build.

OpenCode

  • astronomical launch opencode writes OpenCode's current configuration format and tells OpenCode the model's real context window. astronomical launch codex now declares the context window too.
  • A new OpenCode plugin keeps every model Astronomical serves in OpenCode's model picker: opencode plugin add opencode-astronomical. It is published separately from the app; see the local API page.

Also

  • The model engine is now built as several smaller parts. Nothing changes in what it does.

Your models, chats and settings are kept across the update. Requires an Apple silicon Mac with macOS 14 or later.

Astronomical 0.2.119 is signed with a Developer ID, notarized by Apple, and delivered through the direct download and the signed Sparkle update feed.

Astronomical 0.2.118

Choose a tag to compare

@aosama aosama released this 18 Nov 13:09
ed69875

Faster streaming from the SSD

When a mixture-of-experts model is bigger than the memory you give it, Astronomical reads the experts each word needs from the SSD while it writes. This release reads them with a new native reader, straight into the memory the graphics processor works from, with no second copy and no detour through the macOS file cache.

  • Ornith 1.5 35B-A3B 4-bit as an Expert Atlas, 10 GB memory limit, nothing held in the macOS file cache: 22.9 tokens a second, up from 14.7.
  • The same model at 20 GB: 19.7 ms a token, down from 26.1.
  • Held whole in memory: as fast as before.
  • Replies are word for word the same as before.

New in the catalog

  • Ornith-1.5-35B-A3B-MLX-8bit-vision as an Astro Atlas. The same weights as the 8-bit vision build, stored so each expert is read from the SSD in one step. With a 10 GB limit it writes 16.7 tokens a second against 12.1 for the regular build.

Also

  • astronomical models convert-atlas now converts vision models made with astronomical vision add, whose vision tower sits in its own model-vision.safetensors.
  • Astro's skills each have one consistent icon, and the home cards use the graphite palette.
  • A failing drive is now reported as a drive error, not as a model file that is too short.

Compatibility

Your settings, conversations and downloaded models carry over unchanged. Expert Atlas models made by earlier releases keep working.

Astronomical 0.2.118 is signed with a Developer ID, notarized by Apple, and delivered through the direct download and the signed Sparkle update feed.

Astronomical 0.2.117

Choose a tag to compare

@aosama aosama released this 18 Nov 01:17
a75706e

Fixed

  • Screenshots from your coding tools now reach the model. Agent apps such as GitHub Copilot hand a tool's screenshot back to the model inside the tool's result. Astronomical used to refuse that on the Responses endpoint ("function_call_output is unsupported") and quietly drop the picture on Chat Completions. Now the model sees it, on every endpoint, so a vision model can look at the page your agent just captured. In GitHub Copilot, tick "Supports image input" for the model in Settings → Model providers.
  • A clearer error when an image in a request names a stored file or has no image in it.

Under the hood

  • Test comments brought up to date with the code.

Compatibility

  • Apple silicon Macs with macOS 14 or later.
  • Your settings, models and conversations carry over unchanged.
  • A model that only reads text still declines images, now including images a tool returns.

Astronomical is signed with a Developer ID, notarized by Apple, and delivered as a signed Sparkle update.

Astronomical 0.2.116

Choose a tag to compare

@aosama aosama released this 17 Nov 22:25
ac94231

What's new

  • Faster replies from the Ornith 1.5 35B-A3B model on M5 Macs. When this model comes with its DFlash2 draft model, the draft model guesses a small tree of next words each round, and the large model checks all of them in one step. Replies decode about 1.28 times faster on short chat and about 1.24 times faster after a 10,000-token prompt. Every word is still the large model's own choice, sampled at your settings. From 2,048 tokens of conversation, a new M5 kernel reads the conversation once for the whole tree instead of once per guess, so long chats keep the gain.
  • A new model in the catalog: Ornith-1.5-35B-A3B-OptiQ-DFLASH2-4bit. It is the OptiQ 4/8-bit Ornith 1.5 35B-A3B with its DFlash2 draft model packed inside, about 22.9 GB. Get it from Observatory's Models page, or run astronomical models download Ornith-1.5-35B-A3B-OptiQ-DFLASH2-4bit. On Macs without an M5 GPU it runs exactly like the plain OptiQ build.

Fixed

  • The Tuning page's temperature now takes effect while requests keep arriving. A change to temperature and the other request defaults used to wait for an idle engine, and could be refused while a client kept sending requests. It now applies to the next request, without restarting the engine.

Under the hood

  • Updated crc32fast to 1.5.5 and a CI action to its latest patch.

Compatibility

  • Apple silicon Macs with macOS 14 or later.
  • Your settings, models and conversations carry over unchanged.
  • The faster drafting needs an M5-family GPU. Every other Mac runs every model as before.

Astronomical is signed with a Developer ID, notarized by Apple, and delivered as a signed Sparkle update.

Astronomical 0.2.115

Choose a tag to compare

@aosama aosama released this 17 Nov 11:29
80db12b

What's new

  • Qwen 3.8 27B in the Library. You can now download TensorFold/Qwen3.8-27B-MLX-4bit from the Library or with the astronomical command. It is a large all-round model that reads pictures and text, reasons before it answers, can use tools, and keeps a large memory window for long conversations and documents. It is a 15.0 GiB download. (#2305)

Compatibility

  • Your installed models, conversations and settings carry over unchanged; no settings were added or removed.
  • Requires macOS 14 or later on Apple silicon.

Distribution

This release is signed with a Developer ID certificate, notarized by Apple, and delivered through the signed Sparkle update feed.

Astronomical 0.2.114

Choose a tag to compare

@aosama aosama released this 17 Nov 10:27
1ec4f22

Decision models: train long texts and 9B models as a whole

  • Whole-model training of Clef 9B and Kev 9B now finishes. Since 0.2.112, astronomical decision train --whole-model on a 9B decision model trained and then failed at the very end with "the worker is unavailable", because the trained weights were too large to hand back. They now pass from the engine to the command line through a private temporary file, which is deleted once read.
  • Long texts no longer run out of memory. A text too long to train on all at once is now trained layer by layer: each layer is computed again when it is needed instead of being kept. The trained model is exactly the same, it takes a fraction of the memory, and it is slower; the report says when this happened. On Clef 9B, one long text with ten questions now trains in about five minutes on a 48 GB Mac.

Your existing configuration and trained models keep working as before.

This release is signed with Astronomical's Developer ID, notarized by Apple, and delivered through the signed update feed.