First screen, captured 5 days ago. Open the video.

Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU

What HN thinks

A 284B model squeezed into 64GB at 2 bits, and the thread squints at the readme wondering what got squeezed out.

2-bit quants are lobotomized
Many doubt output quality at such aggressive quantization and call the readme AI slop; some report the tool barely beats plain llama.cpp or fails to run DeepSeek at all.
Local models are good enough
Others argue cheap local rigs now reach useful intelligence, and big models at low quants can still beat smaller models at higher precision.
Frontier subscriptions still win
Heavy agentic users say Opus-class models save far more time than local setups and are cheap relative to their value.

Consensus: Most agree quantization level matters hugely and should have been stated clearly, and that the AI-generated readme is hard to read.

Recommended in the comments

llama.cpp (tool): Commenters suggest using and tuning it directly, noting this project mostly wraps it.
Lemonade llama.cpp ROCm ↗ (tool): Runs a 2-bit DeepSeek Flash at about 15 tok/s on Strix Halo for one user.
gufo ↗ (tool): Recommended as a faster alternative to halogen-server.
Unsloth ↗ (service): Its quant-vs-KLD graphs help judge real quantization damage.
llama-swap (tool): Lets users juggle multiple llama.cpp forks transparently.

Summarised by Claude from the comment thread 5 days ago.

Points: 462
Comments: 200
Best rank: #1
Posted: 2026-11-08 by joshcsimmons
Category: AI & ML

deepseek cpu-inference llama-cpp local-llm quantization mixture-of-experts speculative-decoding

Looks similar

Screenshot of github.com: Show HN: WolveKafka – Consume Kafka with a plain HTTP endpoint

Show HN: WolveKafka – Consume Kafka with a plain HTTP endpoint (github.com)

4 points by piscesdk 3 days ago | 0 comments

GitHub repository for WolveKafka, an open-source Go service that consumes Kafka topics and delivers records to HTTP handlers with retries, DLQ, batching, and offset management.

kafka golang http dead-letter-queue

Screenshot of github.com: Designing Rust APIs around KEM, AEAD and key derivation workflows

Designing Rust APIs around KEM, AEAD and key derivation workflows (github.com)

5 points by lecodev 7 days ago | 0 comments

GitHub repository for NEXUS-Q, a Rust post-quantum cryptographic engine using ML-KEM, ML-DSA, AEAD and key derivation, with multi-language SDK bindings.

post-quantum-cryptography rust ml-kem ml-dsa

Screenshot of github.com: The engineering around post-quantum cryptography is as important

The engineering around post-quantum cryptography is as important (github.com)

3 points by lecodev 9 days ago | 0 comments

GitHub repository for NEXUS-Q, a Rust post-quantum cryptographic engine implementing ML-KEM and ML-DSA with multi-language SDK bindings.

post-quantum-cryptography ml-kem ml-dsa rust

Screenshot of github.com: We've built an open-source feature flag server for 11 years

We've built an open-source feature flag server for 11 years (github.com)

137 points by ivarconr 7 days ago | 97 comments

GitHub repository for Unleash, an open-source (AGPL-3.0) feature management and feature flag platform with SDKs and a web dashboard.

feature-flags unleash feature-toggles continuous-delivery

Screenshot of github.com: Show HN: Write GitHub Actions in any language: Loopt

Show HN: Write GitHub Actions in any language: Loopt (github.com)

7 points by Bnjoroge 4 days ago | 0 comments

GitHub repository for Loopt, a Rust tool that lets you write GitHub Actions workflows in languages like Rust, Python, TypeScript, and Go and compiles them to YAML.

github-actions ci-cd yaml rust