--- title: Show HN: Open-source engine running Gemma 4 26B in 2 GB... description: TurboFieldfare is an open-source Swift + Metal runtime that runs Google's Gemma 4 26B-A4B instruction-tuned model in approximately 2 GB of RAM on any Apple... canonical: https://daily.dev/posts/show-hn-open-source-engine-running-gemma-4-26b-in-2-gb-ram-on-any-m-series-mac-nwy9umvdc twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Show HN: Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Mac | daily.dev og:description: TurboFieldfare is an open-source Swift + Metal runtime that runs Google's Gemma 4 26B-A4B instruction-tuned model in approximately 2 GB of RAM on any Apple... og:url: https://daily.dev/posts/show-hn-open-source-engine-running-gemma-4-26b-in-2-gb-ram-on-any-m-series-mac-nwy9umvdc og:image: https://api.daily.dev/og/posts/nWy9UmvDc.png og:image:alt: Show HN: Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Mac og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Show HN: Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Mac **[Hacker News](https://daily.dev/sources/hn)** · 11 min read · 0 upvotes · 0 comments ## Summary TurboFieldfare is an open-source Swift + Metal runtime that runs Google's Gemma 4 26B-A4B instruction-tuned model in approximately 2 GB of RAM on any Apple Silicon Mac, including 8 GB models. Instead of loading the full 14.3 GB model into memory, it keeps only the 1.35 GB shared core and FP16 KV cache resident, streaming required MoE experts from SSD on demand. The project includes a native SwiftUI Mac app, CLI, streaming model installer, and an OpenAI-compatible loopback server. Benchmarks show 5.1–6.3 tok/s on an 8 GB M2 MacBook Air and 31–35 tok/s on an M5 Pro. The runtime uses custom Metal kernels for quantized GEMV, attention, MoE, normalization, RoPE, and sampling, with 4-bit MLX affine weights and an 8-bit router. Requires macOS 26, Metal 4, and Swift 6.2. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Community take How the wider developer community reacted, aggregated from 1 discussion and 76 comments across hackernews (as of 2026-07-29). **TL;DR:** The community is genuinely impressed by the project's ability to run a 26B MoE model in just 2GB of RAM via SSD streaming and expert caching on Apple Silicon. Most discussion centers on explaining the M2 vs M5 performance gap (SSD speed and RAM caching) and comparing the approach to alternatives like llama.cpp and MLX. **Sentiment:** 72% positive · 20% mixed · 8% skeptical **The case for** - Expert caching achieves ~67% hit rate with 16 slots, meaningfully reducing SSD I/O per token. - Using pread instead of mmap yields a large practical speedup (0.5 tok/s vs 4 tok/s on M2) by overlapping reads with GPU work. - Saves ~12GB of RAM compared to MLX while achieving ~35 tok/s, roughly comparable to ChatGPT response speed. - Reads don't wear out flash memory, so continuous SSD streaming is not a hardware concern. - A macOS 15 workaround was quickly shared by a community member, broadening compatibility. **The pushback** - Apple-only due to Metal and unified memory architecture; Windows would require a completely different approach. - Gemma is not considered the best model for coding tasks; commenters prefer Qwen for that use case. - Performance on 8GB M2 (5–6 tok/s) is marginal for practical use; the big gains require M5 or large RAM. - The approach is architecture-specific and extending it to other models (e.g., Qwen) requires significantly more implementation work. - On a fanless MacBook Air, sustained inference could cause thermal throttling over long runs. **By community** - hackernews (positive): Commenters are enthusiastic and technically engaged, diving deep into SSD caching, memory bandwidth, and comparisons with llama.cpp/MLX, with the author actively responding. **Hottest debate:** Whether the M2 vs M5 performance gap is primarily explained by SSD speed differences or by the larger RAM on the M5 enabling more OS-level file caching. **Open questions** - Can this approach be generalized to other MoE architectures like Qwen or Kimi K3, and what would that require? - How does performance compare directly to llama.cpp with SSD offloading enabled? - What are the spatiotemporal access patterns for a given expert across tokens? - Will a fanless MacBook Air thermally throttle during sustained overnight inference workloads? **Highlights** > My first version used plain `mmap`. On the 8 GB M2, a cold 3.36 MB expert took 10 ms with mmap and 2.8 ms with `pread`. The full simulation was 0.50tok/s for `mmap` vs 4 tok/s for `pread` With `mmap`, OS loads pages reactively as the model touches them. It doesn’t know which experts were selected or when their reads could overlap with GPU work And common weights still use mmap for simplicity So, I believe llama.cpp might run it under 2gb, but I assume it will be slower > — [gitpusher42 on hackernews · 1 comments](https://news.ycombinator.com/item?id=49099602) > The full route changes almost every token. The cache works through partial reuse, about 40% of experts repeat on the next token and 57% within two tokens, cutting I/O from 166 to 88 ms/token on M2 Mac. The longest exact repeat we found was only two tokens. Coding tasks may have higher reuse if code related experts are selected repeatedly > — [gitpusher42 on hackernews](https://news.ycombinator.com/item?id=49099453) > It depends on the use case. I measured this exact model with a 4k context on the mlx engine. It runs at 75 tok/s on my M5 Mac Pro and using 14 GB of RAM. For my engine the same model uses 2 GB of RAM and produces 31–35 tok/s. The project is still experimental so performance may vary as it continues to improve. If you want to save around 12 GB of RAM for other tasks and you are ok with 35 tok/s (afaik it is roughly comparable to ChatGPT’s speed for basic responses) my engine may be a good fit. If you need maximum speed and flexibility just use MLX > — [gitpusher42 on hackernews](https://news.ycombinator.com/item?id=49098926) > With my M1 MBA, I am still on macOS 15. To compile it, just remove the two lines with opts.languageVersion = .version4_0 or surround them with if #available(macOS 26.0, *) { opts.languageVersion = .version4_0 } You'll miss out on a prefill speedup of 2.4x (as it yields 11.24x faster attention), according to the git comments, but it works. (On the 8-GPU-core MBA M1, I get 5.1 tok/s.) > — [xenonite on hackernews · 1 comments](https://news.ycombinator.com/item?id=49099858) > The M5 SSD's performance uplift was fairly substantial, even when compared to the prior generation. > In the Blackmagic Disk Speed Test, the SSD in the M5 MacBook Pro achieved read speeds of up to 6,323 MB/s, compared to just 2,031 MB/s on the M4 MacBook Pro. It's not like the M4 is "slow" in a vacuum, but the M5 SSD is over three times faster, which is a great generation uplift. https://www.tomshardware.com/laptops/macbooks/m5-macbook-pro... > — [GeekyBear on hackernews · 1 comments](https://news.ycombinator.com/item?id=49099000) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=49098510) · 63 points · 76 comments ## Similar posts on daily.dev - [GitHub - mattmireles/gemma-tuner-multimodal: Fine-tune Gemma 4 and 3n with audio, images and text on Apple Silicon, using PyTorch and Metal Performance Shaders.](https://daily.dev/posts/github---mattmireles-gemma-tuner-multimodal-fine-tune-gemma-4-and-3n-with-audio-images-and-text-on-9dwp32aex) · Hacker News · 1 upvotes · 0 comments - [While I slept, my 5-year-old MacBook ran Gemma 4 locally and indexed a year of video — simbastack](https://daily.dev/posts/while-i-slept-my-5-year-old-macbook-ran-gemma-4-locally-and-indexed-a-year-of-video-simbastack-ytj7jpjta) · Hacker News · 0 upvotes · 0 comments --- Tags: [#swift](https://daily.dev/tags/swift), [#gemma](https://daily.dev/tags/gemma) [View this post on daily.dev](https://daily.dev/posts/show-hn-open-source-engine-running-gemma-4-26b-in-2-gb-ram-on-any-m-series-mac-nwy9umvdc) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Show HN: Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Mac","url":"https://daily.dev/posts/show-hn-open-source-engine-running-gemma-4-26b-in-2-gb-ram-on-any-m-series-mac-nwy9umvdc","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/show-hn-open-source-engine-running-gemma-4-26b-in-2-gb-ram-on-any-m-series-mac-nwy9umvdc"},"datePublished":"2026-07-29T15:42:21.393Z","dateModified":"2026-07-29T17:13:43.512Z","description":"TurboFieldfare is an open-source Swift + Metal runtime that runs Google's Gemma 4 26B-A4B instruction-tuned model in approximately 2 GB of RAM on any Apple...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/97c6585327255e1e2dd90ee99a500e75?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/97c6585327255e1e2dd90ee99a500e75?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/show-hn-open-source-engine-running-gemma-4-26b-in-2-gb-ram-on-any-m-series-mac-nwy9umvdc","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"swift,gemma","timeRequired":"PT11M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"Show HN: Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Mac"}]} ```