I shrank DeepSeek V4 Flash to 57GB and it wrote a compiler on my Mac
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…
In plain words
DeepSeek-V4-Flash-0731-Coder is a specialized 57GB model package optimized for coding tasks, built by reducing the original 284B parameter model while preserving reasoning and tool-calling capabilities. It runs on Silicon Macs using the MoEspresso engine and can generate functional code, including compilers. The package is designed for developers and AI agents working on coding tasks, with support for context windows up to 128K tokens and inference speeds around 5 tokens per second on machines with 32GB memory.
written from the facts on this page · September 2026
From the sources
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
This is a 56.869 GB coding-specialized MoEspresso package derived from DeepSeek-V4-Flash-0731 . Its model safetensors occupy 56.831 GB, 32.63% less than the 84.355 GB MoEspresso 2.37 bpw package . The package removes about 80B parameters, leaving 204B total, with about 13B active per token. The package requires MoEspresso 2.1 or newer. MoEspresso is the only engine that serves it. llama.cpp, transformers, and mlx-lm will not load these files. brew install steadfastgaze/tap/moespresso moespresso serve /path/to/package This package is intended for coding and coding-agent workloads. It is not a smaller general-purpose replacement for the 2.37 bpw package. The measurements below show both the…from huggingface.co
In the maker’s words, at launch
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.
Does a similar job
all alternatives →
Distilling DeepSeek into GPT-OSS doesn't transfer censorship. Try itJul 2026 · ctgt.ai · ▲170We recently used DeepSeek V4 Flash as a teacher for finance tasks with GPT-OSS-120B. Distillation works well on this problem. At a constrained 8k token budget, our self-distilled 120B scores 83.61% on FinanceReasoning, above Kimi K3 (81.93%) and Inkling (65.13%). We released the 20B open weights. With V4 as the teacher though, we realized it would be timely to measure if the censorship characteristic of it transferred to the distilled version of the base model. tl;dr it didn't, the teacher answered politically sensitive questions 7 SDs differently than expected, but the distilled model's…
Running 104GB Qwen3.8-Flash-Next on 48GB Mac with at ~12 tok/s5d ago · github.com · ▲239I built slotstream, a way to run Qwen3.8-Flash-Next 4-bit on a low-memory mac starting from 16GB, a 125B parameter model that would need 100GB+ memory/RAM, thanks to expert-offloading/ssd-streaming. Easy to install/update, and mac-native using MLX and Swift. It ships with auto-mode, which makes a good tradeoff between memory usage and speed. I'll be implementing and porting the MTP module for speculative decoding next Local models really are the future of computing!
More dev tools this month
the category →
OpenTrailPaper is open-source bike computer firmware for the LilyGO T5S3 4.7" E-Paper PRO. It supports offline maps, GPX routes, FIT recording and Bluetooth sensors.
Dev tools · 2d ago · opentrailpaper.com
Open-source GTM skills for technical founders
Dev tools · 30d ago · gtmcofounder.com
Launched alongside, August 2026
the whole month →- TL
Life & fun · 11d ago · louisabraham.github.io
Hello HN! I found that picking out plausible but diverse skin tones for my digital art and game development projects was kind of difficult, and I got curious about if there was a way to define a color space that made it easy. I've built a color picker and procedural generation algorithm based on the space as well as a bunch of other fun js features and demos throughout the page that use the equations. If you find it interesting, I have lots of explanations of how I built it and what properties the space has. The methodology might be a bit shaky, but hopefully the result is as helpful for…
Life & fun · Aug 2026 · toneyalexander.github.io
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com