Skip to content

Releases

Spindle v0.9.0

Choose a tag to compare

@ajctools ajctools released this 15 Nov 06:26

Focus gives Chat & vision a quieter place to read and write. Open it from the chat toolbar, double-click the desktop wordmark, or tap the wordmark on mobile.

  • Read opens on the latest formatted reply. Chat brings back the full conversation.
  • Write expands the composer. Hiding it keeps the draft, and sending gives the response the screen again.
  • Live device gauges, run status, and Stop run stay close at hand. Regular desktop chat uses a narrower layout on wide screens.

The source ZIP includes a SHA-256 manifest for each allowlisted file.

Spindle v0.8.1

Choose a tag to compare

@ajctools ajctools released this 14 Nov 19:39
  • Chat formats answers as they stream, with readable inline emphasis, section headings, quotes, and tables.
  • Long structured replies include an In this reply outline. Wide tables and code scroll inside the reply on narrow screens.
  • Copy block preserves the text inside a code fence, including whitespace. Copy reply retains the original Markdown.
  • Saved replies use the same formatting. Model-supplied HTML stays inert and remote images are not embedded.

Use the attached source ZIP to upgrade. Wait for active runs to finish, preserve the existing data directory, restart the helper, and reload Spindle.

Spindle v0.8.0

Choose a tag to compare

@ajctools ajctools released this 13 Nov 22:00

Spindle 0.8.0 makes automatic web search more reliable and adds textured GLB export for compatible Tiiny 3D models.

  • Chat search. Tiiny can turn the current message into a short search query and skip search for local tasks. Spindle now accepts the query format returned by Qwen. If search fails, Chat can still reply and shows why it has no web sources. An explicit query still overrides automatic search planning.
  • Chat stats. Live generation speed appears as an estimate until Tiiny reports its own decode rate. The cache display distinguishes an unavailable measurement from a reported zero.
  • 3D texturing. Choose Apply texture on a compatible saved shape in the 3D workspace or Library & runs. Spindle saves a separate textured GLB. Open the download in a 3D viewer to see the textures.

Download the source ZIP, extract it, and run node server.mjs with Node.js 22 or newer. Open the local Spindle workspace and keep your existing data folder when upgrading.

Automatic search was verified with a live Tiiny and DuckDuckGo. Automated tests cover query handling, search fallback, and the 3D shape-to-texture-to-export flow with a mock device.

The tagged build passed GitHub Actions on Windows, macOS, and Linux.

Spindle v0.7.1

Choose a tag to compare

@ajctools ajctools released this 13 Nov 04:49

Spindle 0.7.1 adds text attachments to Chat and makes dashboard tables sortable.

  • Paste more than 4,000 characters to create a removable attachment, or use Text files to upload UTF-8 files. Read or download sent attachments in Chat and Library & runs. Web search pauses on turns with attached text unless you enter a separate query.
  • Click dashboard column headings to sort or reverse the order. Tiiny’s weekly counter lists apps and models by reported tokens, highest first.
  • Copy prompts, replies, and Meeting Desk transcripts and summaries. Open saved Library results even when the helper has no configured Tiiny connection.
  • This release also includes the compact mobile composer, model retention after successful chat turns, and short native Tiiny clips without FFmpeg. Trimming and stitching still require FFmpeg and FFprobe.

Download the source ZIP, extract it, and run node server.mjs with Node.js 22 or newer. Open the local Spindle workspace. Keep your existing data folder when upgrading.

Validation covered the isolated test suite, Windows/macOS/Linux GitHub Actions, and a browser sweep of all 11 creation modes at five viewport sizes. Model and provider requests used fixtures.

Spindle v0.7.0

Choose a tag to compare

@ajctools ajctools released this 12 Nov 16:00
5cf5401

What’s new

  • Chat shows elapsed time, context used, generation speed, and prompt-cache hits. A visible indicator shows when the selected model is loading.
  • Device groups Models, Connection, and Security & Privacy. The Field guide follows the current workspace, and the docs have refreshed synthetic screenshots.
  • Search Spindle finds local chats, meeting passages, workflows, models, and PromptDex recipes. Chat can attach selected excerpts from Tiiny Vault.
  • Meeting Desk can draft minutes with links back to the transcript. Workflows can export a rough-cut sheet or build an offline Facsimile Book with optional narration.
  • Speech includes a playable synthetic voice reference and a cloning example for compatible downloaded models.

The illustration demo is read-only and blocks live public-source imports. Vault-backed chats keep raw runtime echoes out of saved results. See the security and privacy notes for the local trust boundary, HTTP transport, and remembered-key storage.

Upgrade

Download the versioned source ZIP below. Use Node.js 22 or newer. Wait for any active run to finish before stopping the existing helper. Keep your local data directory and start the updated helper with node server.mjs. The terminal banner and browser footer show v0.7.0.

The allowlisted source archive and GitHub Actions checks passed on Windows, macOS, and Linux. Model and firmware support varies.

Spindle — Live Meeting Desk transcripts

Choose a tag to compare

@ajctools ajctools released this 11 Nov 21:15

Meeting Desk can now show a transcript preview while you record. Turn on Preview transcript while recording and select an ASR model already loaded on Tiiny, or use Load model for preview. Text appears as Tiiny processes roughly 10-second clips. Busy clips can be skipped. Select Transcribe after recording to save a full transcript and see how many audio sections have finished.

The preview uses the speech recognition model on Tiiny and keeps its text in the current browser view. Microphone capture and live text were tested in Brave with MOSS-Transcribe-Diarize while the chat model stayed loaded. Timestamp precision and transcription quality depend on the ASR model and the recording.

Download spindle-source.zip from the latest release. Extract it, install Node.js 22 or newer, and run node server.mjs. Keep your existing private data/ directory when upgrading.

Spindle — PromptDex and Meeting Desk

Choose a tag to compare

@ajctools ajctools released this 11 Nov 20:45

Spindle is an open-source browser workbench for Tiiny. This update adds reusable prompts, model comparisons, and a first Meeting Desk preview.

  • PromptDex has 20 recipes for writing, model checks, vision, and workflows. Compare two downloaded models on the same input and save your review alongside the answers and timings.
  • Meeting Desk records microphone or shared-tab audio and imports clips up to 30 minutes. It queues transcription on Tiiny and pairs the transcript with playback and an editable research rail. Web research sends only the query you submit.
  • Chat and Security & Privacy show what goes to Tiiny and what goes to the search provider. Send to Tiiny only skips web search for one message. A default chat model can load when the device is empty and idle.
  • Rough Cut helps review transcript-based cut suggestions and export a separate edit. The dashboard adds more usage and cost detail.

Download spindle-source.zip from the latest release. Extract it, install Node.js 22 or newer, and run node server.mjs. If upgrading an existing install, keep its private data/ directory. FFmpeg is used by the media editing workflows.

Meeting Desk is an early preview. The capture source depends on browser support. Exact transcript timestamps depend on the transcription model.

Spindle v0.5.2 — compact chat and remembered defaults

Choose a tag to compare

@ajctools ajctools released this 09 Nov 21:39

Spindle v0.5.2 cleans up the chat composer and makes its defaults fit a local inference device.

What's changed

  • Send sits directly beneath the prompt. Chat options starts collapsed and holds response length, thinking, search controls, attachments, examples, Advanced JSON, model switching, and workflow tools.
  • Ample is the default response allowance: up to 32,768 output tokens. For a smaller reported context window, Spindle reduces this to half the window to reserve room for input. This does not measure actual remaining context or override device output limits. Thinking can share the allowance. The ceiling controls device workload, not a paid token balance.
  • Web search starts enabled and remembers your choice across reloads. Search on/off, provider, response allowance, custom value, and streaming preferences are stored in this browser. Prompts, explicit queries, and API keys are not stored in these preferences. Switching search off is remembered too.
  • The search state and allowance remain visible beside Send. The external-query notice stays visible even while options are collapsed.
  • Updated README screenshot, guidance, privacy documentation, and tests. Browser checks verified the compact layout, submission with options collapsed, and preferences surviving reloads. Thirty isolated tests pass locally.

Upgrade and run

Download spindle-source-v0.5.2.zip, extract it, and open a terminal in the spindle folder. Node.js 22 or newer must be installed and available on PATH. No Tiiny SDK or npm package installation is required.

node server.mjs

Stop the old helper before starting the new version, then refresh http://127.0.0.1:8437. An already-open browser tab continues running its old interface until reloaded. Use Ctrl+Shift+R if it still looks old. Find token controls under Chat options.

Read before connecting

WEB SEARCH STARTS ON. The explicit search query, or your current prompt if blank, goes to DuckDuckGo or Brave over HTTPS. Turn it off in Chat options for local-only inference. The choice is remembered in this browser. Conversation history, images, and your Tiiny key are not sent to the search provider. Queries and snippets are saved in local run history. DuckDuckGo needs no paid subscription; optional Brave Search uses a separate API key and its own account terms.

HTTP Tiiny connections remain unencrypted, including API keys, prompts, files, and responses. Use a trusted private network or direct connection.

Remembered Tiiny keys remain plaintext. Leave Remember off for a session-only connection. OS keychain encryption is not implemented. Brave keys remain in helper memory until forgotten or the helper restarts.

Keep the helper local and single-user. Do not expose its port or publish data/, environment files, or CLI configuration.

Search uses snippets rather than full-page fetching and can fail or be rate-limited. Check model-generated citations against the source links. This is a community preview; firmware and model behavior varies. Docker packaging and standalone installers are not included. The screenshot uses a synthetic model and answer with a public live search query, not real device benchmarks. MIT licensed; independent community project, not an official Tiiny product.

Validation

The Windows, macOS, and Linux CI jobs passed all 30 isolated tests and the release-boundary scan: https://github.com/ajctools/spindle/actions/runs/42704487758

Spindle v0.5.1 — visible response-length controls

Choose a tag to compare

@ajctools ajctools released this 09 Nov 21:22

Fixes the hidden 2,048-token chat response cap in v0.5.0.

  • Model default now adds no Spindle token cap. Chat omits max_tokens unless you select a limit or provide one in Advanced JSON. Your device's own runtime defaults, output limits, and context window still apply; this is not unlimited generation.
  • Maximum response tokens is visible beside the thinking controls. Choose model defaults, 4,096 / 8,192 / 16,384 / 32,768 tokens, or a custom positive integer. Explicit selections override Advanced JSON token limits and carry into workflows.
  • Cutoff messages explain the cause. They show the requested cap when known, distinguish uncapped requests, and suggest resending the prompt when thinking consumed the budget before an answer arrived. Reasoning can share the output allowance.
  • Fixed benchmark and capability-check limits remain intentionally small at 128 and 100 tokens.

Upgrade and run

Download spindle-source-v0.5.1.zip, extract it, and open a terminal in the spindle folder. Node.js 22 or newer must be installed and available on PATH. No Tiiny SDK or npm package installation is required.

node server.mjs

Stop the old helper before starting the new one, then refresh http://127.0.0.1:8437. For reasoning-heavy replies, try the 16,384-token preset if the device's default is too small. A larger ceiling permits more work and can increase runtime; it does not force a longer answer.

Privacy and preview limits

HTTP Tiiny connections are unencrypted, including API keys, prompts, files, and responses. Use a trusted private network or direct connection.

Remembered Tiiny keys remain plaintext. Leave Remember off for a session-only connection. OS keychain encryption is not implemented.

Web search is external and off by default. Enabling it sends the explicit query, or current prompt when blank, to DuckDuckGo or Brave over HTTPS. Queries and snippets remain in local run history. Brave credentials are separate and held only in helper memory.

Keep the helper local and single-user. Never expose its port or publish data/, environment files, or CLI configuration.

This is a community preview. Model and firmware behavior varies. Docker packaging and standalone installers are not included. Automated tests use isolated mocks and do not establish compatibility with every device. MIT licensed; independent community project, not an official Tiiny product.

Validation at publication

Windows and Linux CI tests and release-boundary checks passed. The macOS runner remains queued; macOS validation for this patch is pending. See https://github.com/ajctools/spindle/actions/runs/42701165788 for the current result.

Spindle v0.5.0 — chat, thinking, and web sources

Choose a tag to compare

@ajctools ajctools released this 09 Nov 21:10

Spindle v0.5.0 adds streamed chat replies, model-provided reasoning in a collapsible Thinking panel, and optional web search with source links.

What changed

  • Thinking mode and reasoning effort controls follow the selected model's reported capabilities. Returned reasoning stays separate from the answer and from text passed to the next workflow step.
  • Streaming retains final token usage and timings, preserves partial output after failures, and can be disabled for incompatible firmware.
  • Optional DuckDuckGo search needs no key. Brave Search uses a separate key held only in helper memory. Each search supplies up to five snippets with source links, query, provider, and retrieval time.
  • Chat settings and search state appear in performance comparison profiles. Search-enabled execution times include search latency.
  • Updated guidance, security notes, and an illustrated chat walkthrough in the README.

Run it

Install Node.js 22 or newer from https://nodejs.org/ and open a new terminal so node is available on PATH. Extract spindle-source-v0.5.0.zip, open a terminal in its spindle folder, and run:

node server.mjs

Open http://127.0.0.1:8437 and enter your Tiiny endpoint and API key. No npm package or Tiiny SDK installation is required. Existing users should stop the old helper before starting the new version, then refresh the browser.

Read before connecting

HTTP device connections are unencrypted, including the Tiiny API key, prompts, files, and responses. Use a trusted private network or direct connection.

Remembered Tiiny keys are stored in plaintext. Leave Remember off for a session-only connection. OS keychain encryption is not implemented.

Web search is external and off by default. Enabling it sends the explicit query, or current prompt if the query is blank, to DuckDuckGo or Brave over HTTPS. Conversation history, images, and the Tiiny key are not sent to the search provider. Queries and snippets are saved in local run history. Brave keys stay in helper memory until forgotten or the helper restarts.

Keep the helper local and single-user. Do not expose its port to the internet or other users. Do not publish your data/ folder, environment files, or CLI configuration.

DuckDuckGo is best effort and may rate-limit or require browser verification. Failed searches stop the run; Spindle does not bypass verification or silently answer without the requested search evidence. Only snippets are supplied; full result pages are not fetched. Check model-generated citations against the sources.

This remains a community preview. Thinking requires a model/runtime that returns reasoning. Model and firmware support varies. Workflows are sequential; automatic model-selected tool execution, Docker packaging, and standalone installers are not included. Model restoration can fail after a disconnect or shutdown.

The screenshots show synthetic models and responses, not real device benchmarks. Automated tests use isolated mocks; passing them does not establish compatibility with every device or firmware version.

MIT licensed. Independent community project, not an official Tiiny product.