Google Turbo Quant running Locally in Atomic Chat MacBook Air M4 16 GB Model: QWEN3.5-9B Context window: 50000 Summarising 20000 words in just seconds.. You can do 3x larger context window, processing 3x faster than before!

Download for Mac here:

Absolutely mogged by community notes

Lets go!

Keep building!

The first ever wrapper for local models using google's algorithm

and totally FREE for everyone ;)

you're welcome lol

The first ever wrapper for local models using google's algorithm!

Atomic has made something huge. They’re the first team who managed to bring Google TurboQuant into local AI setups and actually put it in people’s hands LFG

Was waiting all day for something like this since the google turboquant post so I wouldn’t have to do it myself thank you lol

That’s seriously impressive Running a powerful model like Qwen3.5-9B locally with a 50K context window on a MacBook Air M4 shows how fast on-device AI is evolving. The ability to summarize 20K+ words in seconds — without relying on the cloud — is a huge win for privacy, speed, and cost efficiency. Atomic Chat is really pushing the boundaries here. If this is already 3x faster with larger context, we’re getting closer to a future where local AI can fully compete with cloud-based setups

Steal someone’s ai project, slap your name on it. Then forget to change the system prompt and get mogged for it Beautiful 2026 slop

These guys are frauds, just a clone of an open source repo, no credit given to actual builders.

@ollama wen?

@ollama It is already inside. Through Atomic Chat, you can use all open-source models, including those from HuggingFace 🤗

Whoaaa!!!!! Cool

Wow! Can I connect to my OpenClaw?

OpenClaw running locally on your own machine very, very soon 👀 but yeah… that’s still a secret 🤫

It’s amazing- super fast, and the web search works really well.

@ollama could you please add this also?

S+ tier slop ✅Pointless AI video ✅Vibecoded slop website ✅AI generated logo (come on) ✅Badly edited sped up video ✅Vibecoded app that copies ChatGPT UI ✅False claims Also why remove Claude from the contributors on GitHub? There is no way 4 *real* people made this.

That’s actually impressive. Running that level of performance locally changes the game faster, private, and no reliance on external APIs.

50k context window on a MacBook Air. no API key. no subscription. no cloud. this is what KV cache compression actually means in practice. the paper was impressive. this is what it looks like when it runs in your hands.

That’s honestly impressive for a local setup.

builders gonna love this 🦞

could we leverage these models within harnesses like Hermes?

@Scobleizer have you seen that?

Very cool

GJ!!!

Amazing!

looks a bit sus ngl

Locked in nd ready

Total fail trying to install this on 2 machines with 2 LLMs helping. ----------- Subject: Cannot connect to local Ollama server — models not detected Environment: Atomic Chat (latest version, macOS) MacBook Air M4, 24GB RAM, macOS Sequoia Ollama v0.9.x running locally on port 11434 Two models available: qwen3:14b and qwen3.5:9b-q8_0 Ollama API is confirmed working: curl → Returns both models correctly via OpenAI-compatible endpoint What I tried and what failed: OpenAI provider with custom base URLSet Base URL to Set API key to ollama Result: Models dropdown shows only hardcoded OpenAI models (gpt-5, gpt-5-mini, gpt-4.5-preview, etc.). It never fetches from the custom base URL. Clicking Refresh does nothing. Restarting the app does nothing. Foundation-models provider — manual model addAdded qwen3:14b manually under Foundation-models It appeared in the Foundation-models settings list Result: Model never appeared in the main chat "Select a model" dropdown. Cannot actually use it for conversations. Typing model name directly in chat selectorTyped qwen3:14b into the "Select a model" search box Result: "No models found for qwen3:14b" Local API Server toggleEnabled "Start server on launch" in settings Result: No change — this seems to be for serving Atomic Chat as an API, not for connecting to external servers. Expected behavior: Atomic Chat should fetch the model list from the custom base URL ( and display the available Ollama models in the chat selector. Actual behavior: The OpenAI provider ignores the custom base URL for model discovery and shows a hardcoded list. Models added manually via Foundation-models don't surface in the chat UI.

This is all free ?

Check this out — you might want to use my quantized Qwen models as the default

wild

@grok 這跟 有什麼關係？

The context window is the sleeper feature. 50K tokens means you can fit entire codebases, legal documents, or research papers. The MacBook as an AI workstation narrative just got more real.

50,000 context window 😬 what are people even doing with these local models? That’s unusable.