frontpage.
newsnewestaskshowjobs

Open Source @Github

fp.

Open in hackernews

Show HN: I shrank DeepSeek V4 Flash to 57GB and it wrote a compiler on my Mac

https://huggingface.co/steadfastgaze/DeepSeek-V4-Flash-0731-Coder-56.8GB-MoEspressoV2
2•hacklas•31m ago
I built a specialized package of DeepSeek V4 Flash 0731 (originally 284B total parameters, 13B active), preserving reasoning, tool calling and coding capabilities:

https://huggingface.co/steadfastgaze/DeepSeek-V4-Flash-0731-...

I let it write a minimal C compiler targeting ARM64, then test the result with Fibonacci and FizzBuzz programs, and it succeeded in less than 1 hour, with the full recording at:

https://youtu.be/XiwSilmV8B0

You can run it on Silicon Macs with my engine https://github.com/steadfastgaze/MoEspresso, while one of the core libraries developed to obtain this result is available at https://github.com/steadfastgaze/mlx-iqk.

The above recording was on a 128GB memory MacBook M3 Max, but you can also run it on 32GB MacBooks with a very usable context (128K tokens) and projected 5 tok/s. I did try it on a fanless 16GB memory MacBook Air M1 (1.39 tok/s), but unfortunately the available context was very small.

How: - First, efficient quantisation: mlx-iqk takes advantage of IQ_K tensor encoding, more efficient than the ones available via llama.cpp or barebones MLX, originally designed by Iwan Kawrakow - I also changed the layout to a k-contiguous one, to make it faster, at least in this Metal setup.

- Second, expert pruning: each of the 40 learned-router layers had 256 experts, and not all of them are equally important for the coding use cases. I removed 80B parameters - this is a known technique called REAP, shared at https://www.cerebras.ai/blog/reap.

- Third: balancing the cheapest IQ1_S_R4 tensor encoding (~1.5 bits per weight), selectively promoting projections to IQ2_KS or IQ2_K where the measured error reduction justified the bytes.

One of the main ideas was not only to save relevant knowledge, but also to not make it forget how to... stop thinking, how to use reasoning. In the first experiments, it would sometimes reason for thousands of tokens without closing its thinking section, or it would go in loops.

Then I solved this by heavily weighting tool-calling traces and structured reasoning in the calibration mix.

Why all AI content sounds the same (and how to fix it)

https://dreasays.substack.com/p/why-all-ai-content-sounds-exactly
1•dundalk03•1m ago•1 comments

Show HN: Flynt.js – 2.2kb, zero-build, CSP-safe reactivity library for MPAs

https://github.com/marsbos/flynt.js
1•bosmarcel•4m ago•0 comments

The Life and Death of Direct File

https://www.ischool.berkeley.edu/research/publications/2026/life-and-death-direct-file
1•ronbenton•6m ago•0 comments

Show HN: I built my 15-year-old game idea in a week with AI

2•ashitlerferad•9m ago•0 comments

Google's new watch tracks insulin resistance to flag diabetes risk

https://www.latimes.com/business/story/2026-08-13/googles-wearables-will-track-insulin-resistance...
2•brandonb•9m ago•0 comments

Quality by John Galsworthy

https://americanliterature.com/author/john-galsworthy/short-story/quality
1•lain98•10m ago•0 comments

Show HN: Meridian, a local-first work journal that drafts project updates

https://github.com/Meridiona/meridian/
1•adithyaharish•11m ago•0 comments

Lenire – Tinnitus Treatment Using Bimodal Stimulation

https://www.lenire.com/
1•Obscurity4340•11m ago•0 comments

Microsoft.Testing.Platform Crash-Resilient TRX

https://medium.com/c-sharp-programming/microsoft-testing-platform-crash-resilient-trx-e62f7cb6e16d
1•sukhpinder0804•12m ago•0 comments

Sectorforth is a 16-bit x86 Forth that fits in a 512-byte boot sector

https://github.com/cesarblum/sectorforth
2•sigalor•15m ago•0 comments

Show HN: Shippin.io – Your SaaS product deserves to be discovered

https://shippin.io
1•jayeshpadhiar•16m ago•1 comments

Gulf of Mexico is a perfect programming language

https://github.com/TodePond/GulfOfMexico
1•xiaoyu2006•17m ago•0 comments

I'm Creating a Minimal Dark Software Factory

https://joeldare.com/creating-a-minimal-dark-factory
1•codazoda•18m ago•0 comments

Atmo.rsvp is a decentralized events calendar built on the AT Protocol

https://portal.eurosky.tech/apps/3moebodfjt2oq
2•doener•19m ago•0 comments

Affix-io/SDK-light – PQC ML-DSA-65 Signing for proofs

https://www.npmjs.com/package/@affix-io/sdk-light
1•affixio•22m ago•0 comments

Show HN: The wanderer – cozy weekend game with no objective

1•freakynit•22m ago•1 comments

Loveyourclanker.org - Patterns of Agent Interaction

https://loveyourclanker.org/
2•javascriptmick•23m ago•2 comments

AI Can Guess Your Location from Social Media Photos with 87% to 91% Accuracy

https://www.macobserver.com/news/ai-can-guess-your-location-from-social-media-photos-with-87-to-9...
2•01-_-•25m ago•0 comments

Madrid hospital creates first bone metamaterial prosthesis

https://www.euronews.com/health/2026/08/15/madrid-hospital-creates-first-bone-metamaterial-prosth...
3•01-_-•26m ago•0 comments

Rcarmo/bun-packrat: Single-file web archive: SQLite-backed, self-contained HTML

https://github.com/rcarmo/bun-packrat
1•rcarmo•26m ago•0 comments

The Moviemaking Exodus from Hollywood

https://www.politico.com/news/magazine/2026/08/16/hollywood-film-production-illinois-tax-credits-...
2•RickJWagner•27m ago•1 comments

Show HN: The Gateless Gate – Ancient Zen text as 50 procedural 3D scenes

https://killedbyapixel.github.io/GatelessGate/
1•KilledByAPixel•28m ago•0 comments

'Fakers' by Rory Cormac Review

https://www.historytoday.com/archive/review/fakers-rory-cormac-review
1•pepys•29m ago•0 comments

Why is a local workout log a $5–16/month subscription?

https://www.gostay.fit
3•justastudent43•30m ago•0 comments

Show HN: I shrank DeepSeek V4 Flash to 57GB and it wrote a compiler on my Mac

https://huggingface.co/steadfastgaze/DeepSeek-V4-Flash-0731-Coder-56.8GB-MoEspressoV2
2•hacklas•31m ago•0 comments

AgentShield – Offline Rust security scanner for AI agent tools and MCP (<50ms)

https://github.com/aiconnai/agentshield
1•aiconnai•31m ago•0 comments

Founder.best launches a weekly newsletter for founders featuring new products

https://www.founder.best
1•anujshashimal12•32m ago•0 comments

Late-night-regrets – piclaw addon for agent self-correction

https://rcarmo.github.io/piclaw-addons/addons/late-night-regrets/
1•rcarmo•32m ago•0 comments

In praise of designer-ish babies

https://www.economist.com/leaders/2026/08/13/in-praise-of-designer-ish-babies
2•Anon84•33m ago•0 comments

Ordinary WiFi can now identify you with near-perfect accuracy

https://www.sciencedaily.com/releases/2026/08/260811052857.htm
11•jonbaer•34m ago•3 comments