frontpage.
newsnewestaskshowjobs

Open Source @Github

fp.

Open in hackernews

Show HN: We quantized Qwen3.6-35B-A3B to 2-bit: 12.3 GB, 225 tok/s on a 4090

https://huggingface.co/EschaLabs/Qwen3.6-35B-A3B-Escha-W2
2•yzh•1h ago

Comments

riknos314•54m ago
I'm missing an accuracy vs bits graph here.

Cool to see this running but if it's at 10%+ loss vs 16bit that's less cool.

Hard to understand that story from what's published here.

yzh•35m ago
Thanks for the comment! Fair point. We do not publish that curve for this model, and the HF model card should say so more clearly.

We did run an internal, unpublished FP16-anchored ladder on Qwen3-8B previously using the same pipeline:

          MATH-500, MBPP+, commonsense-6
2.00 bpw: 68.8, 76.5, 73.1

2.44 bpw: 70.8, 76.5, 74.3

2.75 bpw: 72.8, 78.8, 74.4

3.00 bpw: 74.2, 82.0, 75.1

FP16: 75.6, 84.1, 74.1

At 2 bits, the 8B model is about 9% below FP16 on MATH-500 and MBPP+, but only about 1% below on commonsense. The loss is concentrated in generative math and code, and largely returns as bits increase.

That is one reason this release is a 35B MoE at 2 bits rather than an 8B model: there is more redundancy to exploit. On the 35B, the worst regression is LiveCodeBench at -4.4 points, or -6.6% relative. The other five benchmarks are within run-to-run noise. You can find more details on benchmark results and their harness on our HF model card page.

We do not have a BF16 arm for the 35B because it is about 70 GB and does not fit our hardware (we did this with one 4090, one L40, and tested on a bunch of other RTX GPUs :), so FP8 is the reference.

The 35B bit-width ladder is measured in KL divergence against FP8:

2.09 bpw: 0.1041

3.09 bpw: 0.0921

4.13 bpw: 0.0892

The 2-to-3-bit improvement is statistically significant. The 3-to-4-bit improvement is not.

Test the Email Verification Protocol with an origin trial

https://developer.chrome.com/blog/email-verification-protocol-origin-trial
1•nreece•6m ago•0 comments

How IPTV delivery works, end to end

https://iptvproviders.online/
1•cynthial485•8m ago•0 comments

South Korea's AI Bubble Just Popped [video]

https://www.youtube.com/watch?v=hy90LdpEUvQ
1•Bender•14m ago•0 comments

I Tried to Make AI Writing Sound Human by Banning AI Words Through Logit_bias

https://www.vincentschmalbach.com/make-ai-writing-sound-human-logit-bias/
1•vincent_s•14m ago•0 comments

What's the Fewest People That Could Repopulate Earth? [video]

https://www.youtube.com/watch?v=1qIvFpcGkqc
1•gmays•14m ago•0 comments

South Korea's AI Bubble Just Popped – Now What? [video]

https://www.youtube.com/watch?v=qeXpUZTdvsI
2•Bender•16m ago•0 comments

Show HN: Free Image Merger – combine photos in-browser, nothing uploaded

https://freeimagemerger.com/
1•winsonaibuilder•16m ago•0 comments

Could the Most Radical Plane Design Since the Concorde Take on Boeing?

https://www.wsj.com/business/could-the-most-radical-plane-design-since-the-concorde-take-on-boein...
1•fortran77•18m ago•2 comments

Long-Term Data Storage for Home Users [video]

https://www.youtube.com/watch?v=ulh9z-JluBs
1•haloboy777•19m ago•0 comments

Leopold Aschenbrenner's Situational Awareness seeks capital raise after AI rout

https://www.ft.com/content/280336bf-dbed-405f-b38e-5af644a21549
2•fatcatsbestcats•19m ago•1 comments

Finding a job as a WordPress dev has become impossible

https://twitter.com/DenisBosire/status/2082523094506217765
2•jv22222•20m ago•0 comments

I built a Docker image factory in Go – 300 lines, no daemon, full OCI compliance

https://cheikhhseck.medium.com/building-a-docker-image-factory-from-scratch-in-go-e3c2e7127106
1•cheikhdev•22m ago•0 comments

Cursed Knowledge

https://www.jmduke.com/posts/cursed-knowledge.html
1•twapi•26m ago•0 comments

Surviving Dreams

https://www.survivingdreams.com/reinforcement-dreaming/
1•Aeroi•27m ago•1 comments

Why Standard Displays Often Fail in Specialized Products

https://duobond-display.com/news/custom-displays/506.html
1•Todorov•29m ago•0 comments

Matt Damon's Biceps in That 'Odyssey' Fight Scene? They're Hers

https://www.nytimes.com/2026/07/29/style/devyn-dalton-matt-damon-stuntwoman-odyssey.html
1•wslh•29m ago•1 comments

Interview with a Sysadmin in 2026 [Corporate]

https://www.youtube.com/watch?v=I7IHInNJZvE
1•vismit2000•31m ago•0 comments

"Gravity is worth asking about."

https://unsung.aresluna.org/gravity-is-worth-asking-about/
2•nozzlegear•31m ago•0 comments

I Gave Claude One Tool to Edit, Master and Transcribe All My Audio [video]

https://www.youtube.com/watch?v=xiMh3d97QtQ
1•DanHodgetts•32m ago•0 comments

Recursive Self Improvement for Coding Agents

https://cline.ghost.io/recursive-self-improvement-for-coding-agents/
1•jhack•33m ago•0 comments

Opus Dreams – chats that weren't saved

https://opusdreams.vercel.app/
1•NickNaraghi•39m ago•0 comments

The cost of memory: How organisms balance thinking and sensing

https://phys.org/news/2026-07-memory.html
1•the-mitr•42m ago•0 comments

It is past time to upgrade to post-quantum encryption

https://www.economist.com/leaders/2026/07/29/it-is-past-time-to-upgrade-to-post-quantum-encryption
1•AsyncBanana•43m ago•0 comments

Why the future is local app

https://getapps.cafe/blog/why-the-future-is-local-app
1•knlam•44m ago•0 comments

Studying AI in the Wild: Reflections from the AI Work Research Group (2025)

https://onlinelibrary.wiley.com/doi/10.1111/joms.70021
1•Tomte•52m ago•0 comments

Hiring game developers Triumph Arcade

https://www.triumphlabs.com/
1•elidenslow•53m ago•1 comments

Cross-Vendor Semantic Void Matrix: Zero-Byte Outputs in GPT/Claude/Gemini/Kimi

https://zenodo.org/records/21696066
1•rayanpal_•56m ago•0 comments

Building an S3 client with 20x more throughput and single core [video]

https://www.youtube.com/watch?v=-p-Xm7VEqSk
1•jorangreef•1h ago•0 comments

YKK Sells 10B Zippers a Year. How Did It Get So Big?

https://www.wsj.com/style/fashion/ykk-sells-10-billion-zippers-a-year-how-did-it-get-so-big-253a24a2
1•hbcondo714•1h ago•0 comments

Show HN: We quantized Qwen3.6-35B-A3B to 2-bit: 12.3 GB, 225 tok/s on a 4090

https://huggingface.co/EschaLabs/Qwen3.6-35B-A3B-Escha-W2
2•yzh•1h ago•2 comments