frontpage.
newsnewestaskshowjobs

Open Source @Github

fp.

Open in hackernews

AirLLM 70B inference with single 4GB GPU

https://github.com/lyogavin/airllm
36•Anon84•2h ago

Comments

cpfohl•53m ago
I’m still slightly confused on what this adds.

Let’s say I wanted to run a full size open weight model. I have a 128GB m3 max laptop.

Does this basically load layers in and out on demand? So I still have to download the full model to disk, but the RAM requirements go way down? The readme calls out that one still needs to connect HuggingFace, which leads me to believe that maybe you don’t even need to download the full model?

dofm•38m ago
If you point it at a huggingface model identifier, it will download it, I assume. No way around that.

It reads like it is keeping only the core and the active layer loaded at any one point, and streams layers from disk; there are several other solutions like this and if my understanding is right, this is probably better than an mmap implementation or just streaming experts in.

pvtmert•19m ago
Not an expert in this field, but the "expert" is consisting of multiple layers. To keep it small in terms of memory print, this project streams each layer (dividing even further).

It also requires extra space because of decomposition of the layers. Normally the file format optimized for compute intense workloads. But here the bottleneck is the memory capacity.

Also guessing that you need to be able to hold at least 3-layers at once in the memory, given M x N = R operation, M is the previous layer, N is next, and R is the result. on the next "layer", the R (result) becomes M, gets computed against the next layer, N, yielding the further result R'. And so on, until all layers are processed.

I assume it's horribly slow, but can be put in a non-intrusive background task...

cpfohl•9m ago
Obviously needs downloading eventually :).

It seems like this tool saves on both disk space and RAM, then. Classic trade off: speed vs space.

hna8hjbqzy•47m ago
Same problem, different day
imenani•46m ago
For anyone wondering “how slow is this?”

IIUC, Kimi K3 on RTX 6000 Ada (48GB) takes 292 s/token

https://github.com/lyogavin/airllm/releases/tag/v3.1.0

jackb4040•27m ago
Ahaha thank you, I naively assumed the unlabeled graph in the readme was tps, not spt!
bensyverson•7m ago
I do think these “run a bigger model than will fit in VRAM” projects are necessary steps, but are they functionally useful or helpful to anyone currently? For example, is anyone out there running a big Qwen for coding on a 16-32GB machine with these techniques?
roger_•34m ago
Seeing a lot of these “run 1TB models with 1GB RAM” projects recently. Most seem vibe coded and probably won’t be maintained.

Hoping a winner emerges with some real momentum behind it.

hiramwen•26m ago
You don't really need a maintainer when codex or claude code can set it up for you; thats how I got Trellis2 working on windows and tiny VRAM despite Microsoft recommending you have 24GB VRAM and Linux. Models are pretty disposable now.
classified•20m ago
How a patchwork of Python modules constitutes an API(?) or whatever(?) and how to use it (does anyone?), is beyond me. There is nothing that I would call documentation, let alone concrete usage examples. I know more amusing ways to waste my time. When I want to play with models, I use llama.cpp.
ilaksh•26m ago
I guess the use case is something like: you have a slightly obsolete Mac or PC or a whole bunch of them, and just need to compose one or more convincing spam emails, but it's fine if it takes a full week to do it?
speedgoose•24m ago
And you don’t pay for the electricity.

Critical CVE issued for hallucinated SQLite vulnerability

https://research.jfrog.com/post/sqlite-critical-cves-or-llm-slops/
277•ymir_e•2h ago•82 comments

Don't be a meat proxy

https://gruhn.me/blog/2026-08-03/
1024•ngruhn•7h ago•442 comments

MiniMax H3 Day-0 Support in ComfyUI: Open Weights, Native Audio, and 2K Video

https://blog.comfy.org/p/minimax-h3-day-0-support-in-comfyui
13•vblanco•14m ago•0 comments

Qwen3.8-Max: A New Bar for Coding and Cowork

https://qwen.ai/blog?id=qwen3.8
787•ai2027•11h ago•391 comments

AirLLM 70B inference with single 4GB GPU

https://github.com/lyogavin/airllm
40•Anon84•2h ago•14 comments

Bonsai: Janestreet's UI Library

https://github.com/janestreet/bonsai
123•KolmogorovComp•5h ago•48 comments

Show HN: Nightcrawler – A local AI pentesting agent running on a smartphone

https://github.com/garagehq/nightcrawler/
38•NickySlicks•2h ago•12 comments

The Abandoned Fish Sauce Terrorizing a Small Canadian Town

https://defector.com/abandoned-fish-sauce-canada-interview
40•ohjeez•2d ago•21 comments

Prevent cognitive debt by manually retyping LLM-generated code

https://ankursethi.com/blog/prevent-cognitive-debt-by-manually-retyping-llm-generated-code/
171•mpweiher•4h ago•131 comments

Rust project goals: Immobile types and guaranteed destructors

https://github.com/rust-lang/rust-project-goals/blob/main/src/2026/move-trait.md
135•paavohtl•7h ago•49 comments

What DMARC Protects You From, and What It Does Not

https://senderledger.com/articles/what-dmarc-actually-protects-you-from
53•adulion•4h ago•14 comments

The true power of regular expressions (2012)

https://www.npopov.com/2012/06/15/The-true-power-of-regular-expressions.html
36•uneven9434•5h ago•25 comments

Octane – React's programming model, compiled

https://octanejs.dev
70•nnx•5h ago•22 comments

ICE Collected Nearly 1M People's DNA Last Year–Including Young Children

https://www.wired.com/story/ice-dna-collection-fbi-codis/
140•BlueBerry2001•2h ago•47 comments

Show HN: Isopolis – Isometric pixel map of SF

https://sf.isopolis.city/
266•nuwandavek•13h ago•59 comments

Train Simulator Controller

https://z80.me/blog/tsc-2026-july/
36•austinallegro•3d ago•1 comments

PISIGuard: Protect your personal and sensitive info when you chat with AI

https://github.com/mohamed--abdel-maksoud/pisiguard
14•mohamed_am83•4h ago•10 comments

Characterizing Warp Divergence from Pascal to Blackwell

https://arxiv.org/abs/2607.23402
11•matt_d•3d ago•0 comments

Show HN: We Fixed UniFi's Slow PPPoE Performance with PPPoE Half-Bridge

https://arcbox.dev/blog/unifi-pppoe-half-bridge-acceleration
39•uneven9434•5h ago•14 comments

9front "This Was Supposed to Be Fun" Released

https://9front.org/releases/2026/08/02/0/
43•birdculture•2h ago•12 comments

Why we write our own C and C++ inference engines

https://localai.io/blog/why-we-write-our-own-engines/
85•eatonphil•2d ago•33 comments

Situational Awareness and the Impending Stock Market Volatility

https://www.emergingtrajectories.com/lh/situational-awareness-bigger-picture/
41•cl42•7h ago•27 comments

Show HN: ssh ssh.place

https://ssh.place
141•jeninh•13h ago•80 comments

CP/M-386 – CP/M for 386 protected mode, derived from CP/M‑68K

https://github.com/johnsonjh/cpm386
83•TMWNN•13h ago•38 comments

Why Book Corners won't sync contributions back to OpenStreetMap

https://www.andreagrandi.it/posts/why-book-corners-wont-sync-contributions-back-to-openstreetmap/
125•pizzaiolo•13h ago•81 comments

SwiftUI After 7 Years

https://ykvm.com/2026/07/swiftui-a-story-of-mediocrity/
224•mpweiher•18h ago•215 comments

Show HN: Kakehashi – Experimental userspace to run macOS binaries on Linux ARM

https://github.com/wie-project/kakehashi
237•vlad_kalinkin•21h ago•60 comments

Note-Taking and Personal Knowledge Management

https://unattributed.cc/note-taking-and-personal-knowledge-management
221•surprisetalk•5d ago•78 comments

Developers are attached to tools because tools encode trust

https://stackoverflow.blog/2026/07/29/developers-are-attached-to-tools-because-tools-encode-trust/
231•HieronymusBosch•4d ago•127 comments

Show HN: A Handwritten Blogging Platform

https://handwritten.blog/
144•emilesilvis•3d ago•70 comments