July 29, 2026
Tiny RAM, huge main character energy
Show HN: Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Mac
A giant AI squeezed onto a tiny Mac, and the comments instantly went feral
TLDR: A new open-source Mac tool claims it can run a giant AI model on laptops with just 8 GB of memory by cleverly loading only what it needs. Commenters were impressed, but quickly turned the thread into a debate over speed differences, practical use, and whether this is genius or just a very fancy workaround.
A developer just dropped TurboFieldfare, an open-source tool that lets a huge AI model run on even an 8 GB Apple laptop by leaning on storage instead of stuffing the whole thing into memory. Translation for normal people: this is like getting a grand piano through a studio apartment door by bringing in only the pieces you need at the moment. And yes, the community immediately treated it like a magic trick.
The first wave of replies was pure awe. One commenter just came out with "Wow, amazing!" while another praised the project’s clever "expert caching" like they were handing out medals at the Olympics of nerd ingenuity. But Hacker News being Hacker News, admiration lasted about three seconds before the interrogation began. The biggest hot take? If your computer already has plenty of memory, should you even use this thing at all? That question landed fast, turning the thread from applause into a mini strategy debate over whether this is a breakthrough for budget Macs only or something broader.
Then came the performance detectives. One commenter was visibly suspicious about the massive speed jump between an M2 MacBook Air and a newer M5 Pro, basically asking, "Wait, is the storage really doing all that?" Another went full lab-coat mode, obsessing over how often the AI swaps its little specialist parts in and out. Meanwhile, the funniest moment came from the wonderfully chaotic, "I wonder if i can run this on my MacBook Neo!" — a perfect little joke that gave the whole thread some meme energy. So yes, the software is impressive, but the real show is the comments: half standing ovation, half forensic investigation, all extremely online.
Key Points
- •TurboFieldfare runs the instruction-tuned Gemma 4 26B-A4B model on Apple Silicon Macs in about 2 GB of RAM by keeping a shared core and KV cache in memory and streaming experts from SSD.
- •The installed text-only model occupies about 14.3 GB of storage, while the pinned model download and repack process is about 15 GB on first setup.
- •The project is implemented in Swift and Metal and includes a native Mac app, CLI, decode service, loopback OpenAI-compatible server, and repack installer.
- •The validated target is an 8 GB M2 MacBook Air running macOS 26 with Metal 4, Xcode 26, and Swift 6.2 or newer.
- •Reported decode performance is 5.1–6.3 tok/s on an 8 GB M2 MacBook Air and 31–35 tok/s on a 24 GB M5 Pro.