6/ Two gotchas Thinking can't be turned off. You set how much with a system line like "Reasoning strength: high" (low, medium, high, xhigh). Meta suggests high or xhigh for coding. On llama-server, -c is split across slots. Run out and you get no answer, not an error.

5/ Turn on the drafter for speed The optional DFlash drafter (~1.6 GB) guesses 16 tokens at a time and the main model checks them. Same output, faster. Meta's numbers: RTX 5090 went from 74.9 to 233.4 tok/s, M5 Max from 26.6 to 50.2.

4/ On llama.cpp? Check your build You need release b10353 or newer. Older builds won't even load the files. Always pass --jinja. For images, use llama-mtmd-cli with the separate vision file (mmproj), not plain llama-cli.

3/ Plug it into an agent you already use Ollama has one-line launchers: ollama launch claude --model muse-glimmer The same pattern works for opencode, hermes and openclaw. Handy for private code, since nothing leaves your machine.

2/ Easiest path is Ollama ollama run muse-glimmer That pulls an 18 GB 4-bit build with image input. On Apple silicon, use the MLX tag instead: ollama run muse-glimmer:30b-mlx It's 19 GB and supports DFlash, the speed-up drafter.

1/ Check your hardware first Meta's 4-bit builds target a 24 GB GPU (the ~17 GB build) or 32 GB (the ~20 GB dynamic build, a bit sharper). Full precision wants 64 GB. Meta says the small build loses about 1% on average across 15 benchmarks.

Meta just published the tech report for Muse Glimmer, its 30B open model built for agents that run on your own machine. Apache 2.0, reads images, no API key. Here's how to get it running locally without losing an hour 🧵

6/ Go in with the right expectations The README says it plainly: very early, expect bugs. Big feature contributions aren't being accepted yet. Try it on a side project first, with approvals on, before you point it at code that matters.

5/ Keep it running in the background On Linux or macOS, run t3 service install, then t3 service status to check it. t3 update asks before restarting, because a restart cuts off running agent turns. On a Mac the service stops when you log out, so stay logged in.

4/ Code from your phone The agents keep running on your machine. Your phone is the remote. Simplest path is T3 Connect: on the host, open Settings, Connections, sign in and turn it on. On a CLI host, run t3 connect. The host has to stay on and reachable.

3/ Change the default permissions New threads start on Full access, so commands and file edits run with no prompts. Fix it under Settings, General, New threads, Permissions. Pick Supervised, or Auto, where the provider's review approves routine stuff and asks about the rest.

2/ Install it Mac: brew install --cask t3-code Windows: winget install T3Tools.T3Code Or grab the desktop app from the project's GitHub releases page. If you go the CLI route, run t3 to start the server and open the local web app.

1/ Log in to your agents first T3 Code doesn't come with its own model. It drives the CLIs you already pay for. Set up at least one before you open it: codex login for Codex, claude auth login for Claude Code, agent login for Cursor's CLI.

T3 Code is everywhere tonight. It's a free, open-source app that runs the coding agents already on your machine: Claude Code, Codex, Cursor, Grok Build, OpenCode, Antigravity. Bring your own subscriptions, steer from desktop, web or phone. How to set it up without regrets 🧵

6/ Cut the filler Delete "This commit...", "improves," "enhances," and "various fixes." If the subject could describe any commit, rewrite it. Quick test: could a teammate find this commit with git log --grep a year from now?

5/ Tag the AI, sign as yourself The kernel asks for an "Assisted-by:" line naming the model and any analysis tools. An AI agent must never add Signed-off-by. Only a human can sign off. You reviewed it, so you own it.

4/ Say what you didn't test The same doc tells AI to state plainly what couldn't be built, tested, or reproduced. Put that line in your message too. "Didn't test on Windows" can save a reviewer an hour.

3/ Borrow the Linux kernel's template Its guidance for AI coding assistants asks for a detailed message that covers the problem, the solution, and a Fixes tag pointing to the commit that caused the bug. That works in any repo, not just the kernel.

2/ Give it context, not just the diff Paste the issue, the error, or your notes next to the output of git diff --staged. Then ask: "Write a commit message. Short subject. Body covers the problem, the fix, and anything risky."

1/ Write the why yourself The diff already shows what changed, and AI can read it. What it can't see is why: the bug report, the customer call, the tradeoff you picked. Type that one line first. Let AI build the rest around it.