TurboFieldfare's Growth Play: Ship Exact Numbers, Not Marketing Claims

The Play

TurboFieldfare's Show HN post didn't claim to be "efficient" or "fast." It published numbers a stranger could check.

The project runs Gemma 4 26B-A4B — "26 billion total parameters," "about 3.88B" active per token — in "~2 GB of RAM" on any M-series Mac. The full model is "approximately 14.3 GB." Minimum spec is stated as "8 GB," and the author says it was "validated on M2 MacBook Air" — the cheapest, least powerful machine in the current Apple Silicon lineup, not the flagship.

What TurboFieldfare disclosed that most launch posts wouldn't

Most "Show HN" posts lead with a demo GIF and a vague performance claim. TurboFieldfare did the opposite:

• The exact memory floor: "the shared 1.35 GB core and FP16 KV cache in memory," with the rest streamed from SSD
• The exact minimum hardware, including the cheapest tier: "8 GB" RAM, "validated on M2 MacBook Air"
• Per-device performance, not a single number: "5.1-6.3 tokens/second decode" on the 8 GB M2 Air, "31-35 tokens/second decode" on a 24 GB M5 Pro
• The exact quantization recipe: "MLX affine 4-bit, group 64; 8-bit router; 4-bit shared and routed experts"
• The limitation, stated plainly: "text-only" — no images, audio, video, or tool execution

That combination — floor, ceiling, and named hardware in between — gave Hacker News readers something to check against their own laptop, not just a claim to accept.

Why it worked

The post reached 573 points and 201 comments. Readers didn't just upvote — they argued with the numbers directly. One top comment asked "I'm curious how your project compares to plain mmap? Because llama.cpp will already run 26B in 2GB of RAM if you really want to" — a direct technical challenge that only makes sense because the original post gave a specific, falsifiable number to challenge. Another simply reacted to the scale of the feat: "26B in 2GB is the engineering equivalent of fitting your entire apartment into a storage unit and still having room to pace."

A vaguer post — "our new engine is memory efficient" — doesn't generate that kind of comment. There's nothing concrete enough to disagree with.

The growth play to steal

You don't need a novel inference engine to use this. The transferable move is: replace every qualitative claim in your launch post with the exact number behind it, including the unflattering low end.

• swap "fast" for the actual number, on the actual hardware you tested
• name your cheapest supported configuration, not just your best-case demo machine
• publish the tradeoff or limitation directly instead of letting someone else discover it in the comments
• give readers a number they can personally go verify — that's what turns a passive upvote into an active comment

Bottom line

TurboFieldfare's growth mechanic wasn't the 2 GB number alone — it was pairing that number with the minimum hardware, the per-device speed, and the stated limitations, all in one post. That specificity is what turned a Show HN launch into a 201-comment technical debate, and it's a move any team publishing benchmarks or a product launch can copy directly.

Source: https://github.com/drumih/turbo-fieldfare