Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Mac
Hi HN,I built a specialized inference engine for running 4-bit Gemma 4 26B-A4B-IT on any M-series Mac using about 2 GB of RAM. It is called TurboFieldfare and is written in Swift and Metal.I have always adored on-device AI. It feels like magic that you can run a powerful NN on your Mac or iPhone. So I wanted to push the limits a bit and run a model whose weights don’t fit in memory.The model’s 4-bit quantized weights occupy roughly 14 GB, which makes running it with conventional inference tools almost impossible on an 8 GB or even 16 GB Mac once the OS, applications, and KV cache are included.The trick is to keep the shared part of the model and the KV cache in RAM, then stream only the routed experts needed for each token from SSD. An SSD is way slower than RAM, so the runtime uses a small expert cache and bounded parallel `pread`. While those reads are in flight, the GPU runs the shared part of the layer.I ran more than 100 experiments. Most didn’t work. A few got me here. The experiments are described in the GitHub repo.It currently generates 5–6 tok/s on an 8 GB M2 MacBook Air and 31–35 tok/s on an M5 MacBook Pro.I also added an experimental OpenAI-compatible local server. It supports streaming and tool calls, and reuses one prompt prefix from the KV cache.Try it! The Mac app is easy to install. On the first run, it will download 15 GB of weights from Hugging Face. The model is surprisingly capable.I would love any kind of feedback!
FL score
out of 100
Verdict
medium confidence
Competition
No competitor data yet
Trend
No signal yet
A technical achievement that solves a real constraint (running 26B models on 8GB Macs) but lacks a clear path to sustainable revenue or defensible moat.
The pain
The gap
Build angle
Strengths
- Genuine technical innovation, 100+ experiments show rigor, not luck
- Solves a real hardware constraint that affects millions of M-series Mac owners
- Working product with measurable performance (5-6 tok/s is slow but real)
- Open-source release builds trust and community; lowers barrier to adoption
- OpenAI-compatible API means drop-in compatibility with existing apps
- Addresses growing demand for on-device AI (privacy, latency, cost)
Risks
- Speed is marginal, 5-6 tok/s makes interactive use painful; users will abandon it for cloud APIs
- No business model articulated; open-source alone doesn't sustain a company
- Narrow TAM, only M-series Mac owners; excludes Windows, Linux, mobile
- Commoditization risk, once techniques are public, competitors (Apple, Qualcomm, inference frameworks) will integrate similar optimizations
- Maintenance burden, supporting a specialized inference engine across OS updates is expensive
- Gemma 4 is Google's model; no defensible IP in the weights themselves
- Expert routing adds complexity; harder to support than standard inference
- Market timing unclear, will users accept 5-6 tok/s, or will they wait for faster hardware?
Fly Labs Method
Is the pain real, is there a gap, is it the right time, can one person build it.
Value Equation
Dream outcome and how likely it feels, against the time and effort it costs.
One-Person Business
Curiosity pull, identity fit, and a path from free value to paid for a solo creator.
Viral Frameworks
Hook strength, shareability, and how cheaply it can be tested.
Builder Lens
Evidence the problem exists, timing, defensibility, and a model that fits on a napkin.
Five lenses, one composite. How scoring works
The angle
No market research recorded for this idea yet.
Semble – Code search for agents that uses 98% fewer tokens than grep
Hey HN! We (Stephan and Thomas) recently open-sourced Semble. We kept running into the same problem while using Claude Code on large codebases: when the agent can't find something directly, it falls back to grep, reading full files or launching subagents. This uses a lot of tokens, and often still misses the relevant code. There are existing tools for this, but they were either too slow to index on demand, needed API keys, or had poor retrieval quality.Semble is our solution for this. It combines static Model2Vec embeddings (using our latest static model: potion-code-16M) with BM25, fused via RRF and reranked with code-aware signals. Everything runs on CPU since there's no transformers involved. On our benchmark of ~1250 query/document pairs across 63 repos and 19 languages, it uses 98% fewer tokens than grep+read and reaches 99% of the retrieval quality of a 137M-parameter code-trained transformer, while being ~200x faster.Main features:- Token-efficient: 98% fewer tokens than grep+read- Fast: ~250ms to index a typical repo on our benchmark, ~1.5ms per query on CPU (very large repos may take longer)- Accurate: 0.854 NDCG@10, 99% of the best transformer setup we tested- MCP server: drop-in for Claude Code, Cursor, Codex, OpenCode- Zero config: no API keys, no GPU, no external servicesInstall in Claude Code with: claude mcp add semble -s user -- uvx --from "semble[mcp]" sembleOr check our README for other installation instructions, benchmarks, and methodology:Semble: https://github.com/MinishLab/sembleBenchmarks: https://github.com/MinishLab/semble/tree/main/benchmarksModel: https://huggingface.co/minishlab/potion-code-16MLet us know if you have any feedback or questions!
AI
Postgres extension for BM25 relevance-ranked full-text search
Last summer we faced a conundrum at my company, Tiger Data, a Postgres cloud vendor whose main business is in timeseries data. We were trying to grow our business towards emerging AI-centric workloads and wanted to provide a state-of-the-art hybrid search stack in Postgres. We'd already built pgvectorscale in house with the goal of scaling semantic search beyond pgvector's main memory limitations. We just needed a scalable ranked keyword search solution too.The problem: core Postgres doesn't provide this; the leading Postgres BM25 extension, ParadeDB, is guarded behind AGPL; developing our own extension appeared daunting. We'd need a small team of sharp engineers and 6-12 months, I figured. And we'd probably still fall short of the performance of a mature system like Parade/Tantivy.Or would we? I'd be experimenting long enough with AI-boosted development at that point to realize that with the latest tools (Claude Code + Opus) and an experienced hand (I've been working in database systems internals for 25 years now), the old time estimates pretty much go out the window.I told our CTO I thought I could solo the project in one quarter. This raised some eyebrows.It did take a little more time than that (two quarters), and we got some real help from the community (amazing!) after open-sourcing the pre-release. But I'm thrilled/exhausted today to share that pg_textsearch v1.0 is freely available via open source (Postgres license), on Tiger Data cloud, and hopefully soon, a hyperscalar near you:https://github.com/timescale/pg_textsearchIn the blog post accompanying the release, I overview the architecture and present benchmark results using MS-MARCO. To my surprise, we were not only able to meet Parade/Tantivy's query performance, but exceed it substantially, measuring a 4.7x advantage on query throughput at scale:https://www.tigerdata.com/blog/pg-textsearch-bm25-fu
AI
Consus (YC F24) – LLM gateway for companies handling sensitive data
Hey Hacker News! I'm Eric, the founder of Consus. We route all AI requests to ITAR restricted and controlled unclassified information (CUI) compliant infrastructure. Consus itself satisfies ITAR restrictions, and we recently completed our FedRAMP equivalency audit against the 410 FedRAMP High controls. FedRAMP High is the U.S. government’s strictest security baseline for cloud services handling sensitive unclassified data. A breach of this data could cause severe or catastrophic harm. Achieving FedRAMP equivalency, or authorization, is required before processing CUI information.Demo: https://youtu.be/6MYxJe6P4oYAerospace and defense is experiencing rapid growth. Unfortunately, most companies are struggling to use AI because of necessary data protections & compliance risk. These companies can move even faster if they could use the AI tools that most other industries take for granted.I spent most of my career prior to Consus in aerospace & defense. I launched national security payloads at SpaceX, and was an early employee at Castelion. I’ve watched first hand how compliance red-tape makes it difficult for companies to adopt new technologies and trust vendors that touch their sensitive data. For example, I worked with a satellite manufacturer that doesn’t allow any AI tools, even for non-sensitive work.Even as AI has become extremely useful over the past 9-12 months, I was surprised by how many aerospace & defense companies refused to touch it. They could see how it could be helpful, but they were scared about model providers training on their data, violating an ITAR rule, or risk losing new business because of a failed cyber assessment. These are all rational fears.There were a lot of ways to solve some/all of these problems, but we focused on building a compliant AI inference layer that can "plug" into any AI tool or workflow. Just as OpenRouter integrates with most tools, Consus can connect to tools like ChatGPT or Clau
AI
Questions about this idea?
FlyBot reads the scoring and gives you a second opinion on “Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Mac”.