distrocool Submit a tool
DistroCool Brief

What changed in AI. What you can build with it.

RSS ↗
← All editions

7 min read

Local AI you can hear, inspect and simulate

Build one small thing, not an all-powerful agent. This edition offers three starting points: give an app a local voice, inspect how a model chooses between fixed options, or explore a physics scene before training a robot. The projects are free to try on supported hardware; the Mac-only requirements and model licenses matter. These are source-checked instructions, not our own performance tests. Pick the experiment that fits the computer you already have.

Make something speak

Give a tiny app a voice without a cloud speech bill

What changed

Gradium released the open-weight Phonon family on October 20; its November 5 follow-up explains the integrations. The English 31M model has Apache-2.0 weights and can generate speech locally through the ptts runtime. This is an existing release you can build with, not a launch today. Start with a single spoken status message rather than a voice-cloning project.

Why it matters

An offline timer, accessibility experiment or game character can speak without sending each sentence to a hosted speech API. The small English checkpoint lowers the download barrier. Local generation is useful for privacy and offline use, but first-time package, model and voice downloads still need a connection.

Try it

With uv installed, open a new scratch folder and run the model card's English example: uvx ptts --lang en --quant q8 --voice Easton "Hello from Phonon." -o speech.wav Play speech.wav and check whether every word is understandable. Then replace only the quoted sentence with one short, non-sensitive status message. If you prefer no installation, try the embedded text-to-speech demo in Gradium's announcement. Do not clone another person's voice for this first attempt.

Availability & cost: The 31M English q8 weights are a public 64.5 MB download, plus runtime and voice files. The runtime offers desktop/Python and browser integrations; the documented Python wheels require CPython 3.9+. The English weights are Apache-2.0; runtime code is MIT OR Apache-2.0. No paid API key is needed for the built-in-voice example.

Limits: The 40M models have different, non-commercial CC BY-NC 4.0 terms and gated downloads; do not assume the English model's license covers them. Device/browser support varies. We verified the documentation, not audio quality, offline caching or latency on your machine. Listen to numbers and names rather than treating generated speech as automatically correct.

Original sources

Inspect an agent decision

Build a ticket router you can question, not just trust

What changed

On October 29, Mattt introduced AnyDecisionModel alongside AnySpeechRecognizer. The Swift decision package returns typed yes/no probabilities, choices and scores instead of requiring you to parse a chat response. Its public examples include a support-ticket router and local model evaluations. Start with the router, not an automated customer-support system.

Why it matters

When an app needs one of three labels, a paragraph is an awkward interface. Seeing each option's probability makes uncertainty easier to inspect and gives you somewhere to put a human-review branch. The package explicitly warns that a confident result can still be wrong; typed output is not a security boundary or proof of correctness.

Try it

On a supported Apple-silicon Mac, open the AnyDecisionModel-Examples repository and inspect its README and mise.toml before trusting its tasks. With Xcode and the Metal toolchain ready, follow the documented sequence: mise trust, then mise run setup, then mise run example. The default route uses local MLX, not the optional Jev API. Read the four sample ticket decisions. Next, try one fictional mixed billing/login ticket and decide where a person should review it. Do not connect a real inbox or automatically forward messages.

Availability & cost: Package and examples are Apache-2.0. The example project requires an Apple-silicon Mac, macOS 26+, Xcode 27 with Swift 6.4+, the Metal toolchain and mise. Its default Qwen3-4B 4-bit model downloads about 2.3 GB on first use. No Jev key or paid inference service is required for that default example; model terms are separate from package terms.

Limits: This is not a Windows quickstart. Setup does not download the model; the first inference does. Confidence describes concentration of the choices, not the chance of being correct. The authors say some evaluation labels have not had human review. We read the code documentation but did not compile or benchmark the Mac example. Never use the demo alone to make consequential decisions.

Original sources

Explore robot simulation

Open a tiny physics lab before buying a robot

What changed

MuJoCo 4.0.0 was published on November 6, with the maintainer's overview following on November 7. The release adds speculative contacts for fast-moving objects and plastic deformation, alongside the new Studio tooling. The project also exposes interactive browser showcases, so a first look does not require a local simulator install or a robot.

Why it matters

Robot-learning software needs an environment to act in, not just a language model. A simulated scene lets you inspect motion and contact behavior before thinking about a policy or training run. MuJoCo is the physics engine, not an AI that learns a task for you; attractive animation is not evidence that a robot will work in the real world.

Try it

Open mujoco.org from the maintainer's announcement and explore one built-in 4.0 showcase; live.mujoco.org is the linked browser model viewer. Spend five minutes observing one scene and write down the initial state, what moves, and one failure you would want to detect. Do not begin a training run. For a later local experiment, the Python docs distinguish mujoco-core for minimal headless simulation from the full mujoco package. Start in a fresh environment rather than upgrading a working robotics project in place.

Availability & cost: MuJoCo's engine source is Apache-2.0 and the browser showcase is publicly accessible without buying a robot or model API. Local use needs a supported Python/platform combination; the full package adds Studio/rendering components, while mujoco-core depends only on NumPy. Training agents and rendering large scenes can require substantially more compute than viewing a demo.

Limits: Version 4 includes breaking changes; existing 3.x Python installations need the documented migration, not a casual in-place update. Sample model scenes are not bundled in the PyPI wheels. Browser/WebGL support and download time vary. We inspected the release and docs, not simulation accuracy, real-world transfer or training performance.

Original sources

4 more signals

Useful checks and updates. Guidance shows its check date; event dates identify actual announcements.

  • Small local model, specific limits: LFM3

    Liquid AI's October 27 LFM3 release includes a 1.7B text model with local GGUF/MLX options. Its model card says not to use it for complex multi-step reasoning or agentic coding. Try a small fictional labeling task, not your whole codebase. The custom LFM Open License has a commercial revenue threshold; open weights do not mean Apache-2.0. Check your runtime's hardware support before downloading.

  • Transcribe your own short clip on an Apple device

    AnySpeechRecognizer lets Swift apps choose recognition engines behind a familiar speech API. Start with a recording you own, not a private meeting. The package is Apache-2.0, but downloaded model licenses differ. Parakeet needs Apple silicon and roughly 632 MB of model files; live mode adds another encoder. Qwen3-ASR has different timing and memory limits. This is a documented path, not our own transcription test.

  • Clean a document before handing it to an AI

    Microsoft's MIT-licensed MarkItDown remains a useful local preparation tool, not a new launch. Convert one non-sensitive text-based document and compare headings and tables with the original before using the Markdown as context. Install only the format extras you need. Image captioning, speech transcription and Azure integrations can use hosted services; do not mistake those optional paths for offline, free conversion.

  • Read the grader before believing a security score

    CWE-bench v2's November 3 write-up describes 100 public defensive coding tasks and programmatic verifiers. A useful free first step is to read what those checks measure, not repeat a leaderboard number as a product guarantee. Downloading the dataset requires accepting its access conditions; running agents needs isolated environments and may incur model costs. We have not reproduced its scores.

Build something cool

Make a one-feature prototype with five honest checks

Pick only one experiment above. Use the free operating-brief workflow to state its input, allowed actions and stop conditions. Save five tiny fictional test cases and expected outcomes: for speech, include a date and a number; for routing, include an ambiguous ticket; for simulation, describe one scene and a failure to watch for. Record what you actually ran, dependencies downloaded and anything that failed. Stop before sending messages, using private customer data or operating real hardware.

Open the DistroCool workflow →

Was this edition useful?

Your choice helps us shape future editions. No account needed.