TurboFieldfare, an open-source runtime released July 29, 2026, runs Google's Gemma 4 26B model in about 2GB of RAM on any Apple Silicon Mac, free and offline. Scored against Claude Sonnet 5 ($2/$10 per million tokens through August) and ChatGPT's GPT-5.6 using Radically Human's IDEAS framework, local AI wins on data sovereignty and cost at scale, but loses badly on reasoning depth for anything a founder would actually pay to get right.
Three days ago a solo developer named Andrey Mikhaylov posted a project called TurboFieldfare to Hacker News and it hit 911 points before I'd finished my coffee. The pitch: a custom Swift and Metal runtime that runs Google's instruction-tuned Gemma 4 26B model, 26 billion parameters, in roughly 2GB of RAM, on an 8GB M2 MacBook Air, the cheapest Mac Apple currently sells. I've had three separate founder clients ask me the same question since: does this mean I can stop paying for Claude or ChatGPT? I spent the weekend actually testing it, and the honest answer is more useful than yes or no.
TurboFieldfare doesn't load the full 14.3GB model into memory. It keeps a 1.35GB shared core resident and streams only the specific experts each token needs from SSD, using 4-bit MLX quantization. On the base M2 Air it decodes at 5.1 to 6.3 tokens per second; on an M5 Pro it climbs to 31-35 tokens per second. It's Apache 2.0 licensed, free, text-only, and requires no internet connection after the initial 15GB download. Compare that to Claude Sonnet 5, priced at $2 per million input tokens and $10 per million output tokens through August 31, 2026 (rising to $3/$15 in September), with a 1 million token context window, or ChatGPT's current flagship GPT-5.6 Sol at $5/$30 per million tokens, with the free tier capped at roughly 10 messages per 5-hour window on a 16K-token model called Terra.
Paul Daugherty and H. James Wilson's Radically Human gives me the actual framework I use to judge this kind of trade-off, because their IDEAS model, Intelligence, Data, Expertise, Architecture, Strategy, was built for exactly this question: when does a smaller, more constrained AI system beat a bigger one? Their example is Mazda improving engine calibration using roughly 1,000 times less data than a conventional deep-learning approach, by trading raw scale for a tighter, purpose-built model. TurboFieldfare is the personal-computing version of that same trade.
| Criterion (IDEAS) | TurboFieldfare + Gemma 4 | Claude Sonnet 5 | ChatGPT GPT-5.6 |
|---|---|---|---|
| Intelligence (reasoning depth) | 2/5 — capable but noticeably shallower on multi-step reasoning | 5/5 | 5/5 |
| Data (privacy/sovereignty) | 5/5 — nothing leaves the machine | 2/5 — cloud, enterprise DPA required for real privacy | 2/5 — same |
| Expertise (fine-tune/control) | 4/5 — open weights, fully local control | 2/5 — prompt-level only | 2/5 — prompt-level only |
| Architecture (integration effort) | 2/5 — you own the ops burden | 5/5 — mature API, SDKs | 5/5 — mature API, SDKs |
| Strategy (cost at scale) | 5/5 — fixed hardware cost, zero marginal | 3/5 — scales with usage | 3/5 — scales with usage |
What that table tells me, after actually running both side by side on real founder work this week: TurboFieldfare is not a Claude or ChatGPT replacement for anything I'd call thinking-partner work, drafting a board memo, red-teaming a pricing decision, synthesizing a messy set of customer interviews. It got noticeably shallower and more repetitive past three or four reasoning steps, exactly what you'd expect from a 26B model against Sonnet 5 or GPT-5.6 Sol's frontier scale. Where it earned its place on my machine is the opposite case: anything I want processed that I don't want touching a cloud API at all, cap table scenarios before a raise, a draft termination letter, an early product idea I'm not ready to expose to any vendor's training pipeline. Radically Human's whole argument is that the winning move isn't always more data or a bigger model, it's matching the tool's constraints to the actual job, and running two systems side by side has made that concrete for me in a way the book's Mazda anecdote never quite did on its own.
The honest limit: this only works on Apple Silicon, it's arm64-only, and the project explicitly warns it can still repeat itself or answer wrong, so I check anything consequential regardless of which system produced it. It's also a hobby project from one developer, not a company with a support line, so I wouldn't bet a client deliverable on it without a fallback.
Sources: TurboFieldfare GitHub repository and README, accessed August 2026; Hacker News discussion, July 29, 2026; Anthropic, Claude Sonnet 5 pricing announcement; BenchLM, OpenAI API pricing, August 2026; Paul R. Daugherty & H. James Wilson, Radically Human (Harvard Business Review Press, 2022).
Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · AI Coach App — Building It in 8 Hours