frontpage.
newsnewestaskshowjobs

Open Source @Github

fp.

Open in hackernews

I pushed Qwen3.8-27B to 99 tps on a single request and 1150 on batch on a 3090

https://github.com/syv-ai/qwen38-27b-rtx3090
1•mess1337•42m ago

Comments

mess1337•42m ago
Yesterday I released the first version of hyper-optimized Qwen3.8-27B inference engine for a RTX 3090, reaching 82 tps on single request and 672 peak.

Over the last 24 hours I've been exploring how to improve it further (without a quality degredation).

Yesterday we had these 3 optimizations: - fp8 kv cache

- lm_head int8

- embed_tokens in8

Now now I added: - fp16 recurrent state

- int8 activations, MLP

- int8 activations on everything

- draft_sample_method=probablistic which samples from the MTP distribution rather than taking the argmax) - does not work for greedy sampling.

For single user mode, the user can enable ctx=fast, which uses FlashAttention, bf16 kv, 64k context and gives the user 99 tps, but at the tradeoff of less context.

Also the prefill is now up to 1812 tps at 1k and 1000 tps at 100k context - an increase of about 50% to 25%.

I've updated the repo will everything you need to start flying: https://github.com/syv-ai/qwen38-27b-rtx3090

Israel creates fake think tank in likely attempt to dupe AI chatbots

https://responsiblestatecraft.org/israel-influence-chatgpt/
1•DeepLogin•1m ago•0 comments

Show HN: Openleetcode – local LeetCode runner where tests live in the repo

https://github.com/therepanic/openleetcode
1•therepanic•3m ago•0 comments

Pure Pleasure Isn't What You Want, Or, Notes on the Jhānas (2022)

https://sashachapin.substack.com/p/pure-pleasure-isnt-what-you-want
1•eatitraw•6m ago•0 comments

Show HN: Engelbart – Mange Goals and TODOs for Claude Code

https://github.com/divadbaroon/claude-plugins
1•hudsonmp•6m ago•0 comments

Notes on purposeful skill maintenance/improvement/neglect

https://www.alexejgossmann.com/skill-maintenance-improvement-neglect/
1•funnygiraffe•7m ago•0 comments

Keep the monolith, but split the workloads (2023)

https://incident.io/blog/monolith
1•mooreds•8m ago•0 comments

You can just choose how many bugs you want now

https://nolanlawson.com/2026/08/16/you-can-just-choose-how-many-bugs-you-want-now/
2•birdculture•9m ago•0 comments

Top Russian economist replaced after warning of economic costs of Ukraine war

https://www.cnn.com/2026/08/17/europe/andrey-klepach-top-russian-economist-replaced-intl
3•mooreds•10m ago•1 comments

Alberta's AI data centre gamble: Big investment or local burden?

https://financialpost.com/technology/alberta-ai-data-centre-investment-local-burden
3•jethronethro•11m ago•0 comments

Community Cat Protocol: Eartipping

https://www.alleycat.org/resources/feral-cat-protocol-eartipping/
2•mooreds•11m ago•0 comments

Show HN: Slaunt – control what AI agents can access, do, and execute

https://slaunt.ai
1•Kerode0•14m ago•0 comments

Show HN: Gene Inspector Pro – I built a tool for exploring your own genome

https://gene-inspector.pro/
1•sergey-a•16m ago•0 comments

Offloading Rust to GPUs Proves Capable of High Performance with Memory Safety

https://www.phoronix.com/news/LLVM-Offload-Rust-Performance
1•speckx•16m ago•0 comments

Ask HN: AI-Haters and AI-Usage

1•emn4tor•16m ago•0 comments

Tell HN: I never mean to flag anything

1•rgbrgb•17m ago•0 comments

YouTube is changing how it counts views to give the numbers a boost

https://www.theverge.com/streaming/981105/youtube-video-view-counting-update
1•01-_-•18m ago•0 comments

Plotly Studio Embedded

https://plotly.com/blog/plotly-studio-embedded-ai-chat-dash-apps/
1•chriddyp•19m ago•0 comments

Freestyle Linked Lists Tricks

https://nullprogram.com/blog/2025/12/31/
1•atiedebee•21m ago•0 comments

GitHub's Recent Crisis Has a Simple Fix: Intelligent Backoff Queues

https://www.birkey.co/2026-08-17-githubs-recent-crisis-has-a-simple-fix.html
1•msolli•21m ago•0 comments

Nvidia backing $105B in financing for OpenAI data center in Ohio

https://www.cnbc.com/2026/08/17/nvidia-financing-open-ai-data-center-ohio.html
1•kristianp•21m ago•0 comments

Should You Use AI for a Task? Here's a Simple Way to Decide

https://www.schneier.com/blog/archives/2026/07/should-you-use-ai-for-a-task-heres-a-simple-way-to...
1•amai•21m ago•0 comments

The 'real Indiana Jones' who found an ancient, lost city of the Incas

https://www.bbc.com/culture/article/20260814-how-a-real-indiana-jones-claimed-he-had-found-the-an...
2•plaguna•22m ago•0 comments

iOS/iPadOS 26.6.1 Security Update

https://support.apple.com/en-us/148282
1•qaz_plm•24m ago•0 comments

What do we know about the effectiveness of digital advertising? (2019)

https://thecorrespondent.com/100/the-new-dot-com-bubble-is-here-its-called-online-advertising
1•hackthemack•24m ago•0 comments

Netanyahu puts NYC mayor Mamdani alongside Iran's Ayatollah on enemies poster

https://www.independent.co.uk/news/world/middle-east/mamdani-netanyahu-poster-israel-b3034086.html
4•handfuloflight•25m ago•0 comments

Reflections on My Work Trial

https://sageox.ai/blog/reflections-on-my-work-trial
1•skadamat•27m ago•0 comments

Elon Musk's GitHub alternative couldn't have arrived at a worse time for GitHub

https://www.neowin.net/news/elon-musks-github-alternative-couldnt-have-arrived-at-a-worse-time-fo...
4•bundie•27m ago•0 comments

Data Loading for AI/ML: A Comprehensive Guide

https://www.lancedb.com/blog/data-loading-guide
3•westonpace•30m ago•0 comments

Cognitive Convergence: Deep Similarities Between LLMs and Human Cognition (2026)

https://arxiv.org/abs/2607.26179
2•eh_why_not•30m ago•0 comments

The Hidden Dangers of Buying Supplements from Amazon

https://www.mensfitness.com/news/dan-go-amazon-supplement-scam
2•bushwart•32m ago•0 comments