GPT-QModel, QvQ ultra low bit quants. Contributor to stuff you are probably using like sglang, vllm, transformers. ucla alum
Earth
- Tweets 7,665
- Following 4,535
- Followers 1,736
- Likes 3,666
You need to have extreme curiousity if you want humans to sustain this war with AI. Either out monkey the AI, or retire. That's my motto. I am not retiring, yet. Gonna out-curious the Ai until that day comes.
Curiosity is my #1 weapon. Not lemming is my #2.
6
1,348
MiniMax M3.1 flash vs DeepSeek v4.1 flash 1017 has no data for many tests. Can you guess why? 🐋
MiniMax M3.1 Flash: Space Performance + Bunny Speed
🚀🐰 Yes, Space Bunny - which has topped global leaderboards with 200T+ tokens - was an earlier version of M3.1 Flash.
💰Redefining the performance economics of agentic AI. M3.1 Flash achieves frontier-level intelligence across key tasks such as Cowork (Automation-Bench: 62%; ALE: 30%), Coding (DeepSWE: 74%) and Visual Reasoning (BabyVision: 91%), at 1/3 the cost of M3 and 3x the speed.
API pricing (flat rate): $0.15 / $0.60 per 1M input/output tokens, with cache reads at $0.015 / 1M tokens.
Model weights will arrive on Oct 23.
API: platform.minimax.io
M Plan: minimax.io/m-plan
MiniMax Code for desktop: agent.minimax.io/download
118
QUIC is a protocol hack around latency issue and introduces 1000x increase in code/buffering complexity. Don't use it unless you have to.
Sorry everyone... We're announcing a 0-Day vuln in QUIC that allows anyone to remotely turn off a server running these libraries. I swear we didn't plan this.
foxio.io/blog/header-accordi…
132
Maybe gitlab should've used Pcre2. Fafo. Btw, I have the fastest python binding for pcre2 in PyPcre. So much faster than anything Python has and it re-uses system libs for instant cve mitigation pushed by OS vendor!
🚨 CRITICAL: GitLab flaw (CVSS 10.0) affects 12 years of releases.
CVE-2026-110264 could allow unauthenticated remote code execution on self-managed servers.
Read - thehackernews.com/2026/10/cr…
1
338
Why in the world aren't more projects using Pcre2? It's battle tested to the max and offers 10x the features other engine does. There is nothing better. "It needs to be ruby/golang native!" Nothing needs to be "native" if it sucks!
github.com/modelCloud/PyPcre
53
I am not using full dims (max bw) on zen5 and have not yet validated on Granite Rapids. I literally can't afford ram for my mobos. Got cpu/mobo with no ram! If you like my work, donate some spare change via x money so I can populate those empty mobos with dims.
github.com/ModelCloud/GPTQMo…
39
YOCO prefill optimization from @deepseek_ai gave my inf/hw stack 15-33% prefill uplift in DeepSeek v4.1 Flash 1017 at various batch/chunks. Now over 10.2k t/s from 9k. Depending on gpus, the gains maybe greater!
2
10
845
Why is HF Kernels so behind in build compilation. It has been so bad, since inception, I only port code from it instead of consuming it directly.
> HF CPU kernel is a standalone GPTQ GEMV/GEMM (gemm_int4_forward) with builds only through Torch 2.12
@art_zucker
2
2
417
Yes! I was hoping this will happen and now it will. Onednn is a very good sdk and I am glad AMD has finally joined.
Intel and AMD are working together to simplify AI development on x86.
By bringing @PyTorch CPU optimizations together through oneDNN, we’re creating a more consistent software path for developers across Intel and AMD processors.
Learn more: ms.spr.ly/6011avvs3
3
400
I am stunned by this announcement. There is a strategy here from the DeepSeek team that I am not fully understanding yet. They are figuratively taking out their competitors one by one, using opensource.
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
2
221
Better than DS v4.1 Flash but not 1017 edition.
MiniMax M3.1 Flash: Space Performance + Bunny Speed
🚀🐰 Yes, Space Bunny - which has topped global leaderboards with 200T+ tokens - was an earlier version of M3.1 Flash.
💰Redefining the performance economics of agentic AI. M3.1 Flash achieves frontier-level intelligence across key tasks such as Cowork (Automation-Bench: 62%; ALE: 30%), Coding (DeepSWE: 74%) and Visual Reasoning (BabyVision: 91%), at 1/3 the cost of M3 and 3x the speed.
API pricing (flat rate): $0.15 / $0.60 per 1M input/output tokens, with cache reads at $0.015 / 1M tokens.
Model weights will arrive on Oct 23.
API: platform.minimax.io
M Plan: minimax.io/m-plan
MiniMax Code for desktop: agent.minimax.io/download
136
FloatDoor. Hilariously diabolical. We are seriously going to see stuxnet like trojan horses soon.
arxiv.org/abs/2606.19535 They backdoored an LLM so that it writes insecure code when runing on specific GPUs!
1
252
👀👀👀
Quantinuum has achieved a major industry breakthrough that solves a critical error-correction bottleneck and validates our path to delivering a fault-tolerant quantum computer by 2029.
We’ve announced a new world record for quantum error correction, achieving 99.9995% logical fidelity on our Helios commercial quantum computer using our error-correction architecture.
Equivalent to just five errors in a million operations, the result is 10× better than our previous record achieved just last month and 100× better than the best published result on any other hardware modality.
The study also demonstrated a record 12× reduction in logical errors as code distance increased, a critical step toward scalable fault tolerance.
These successive breakthroughs using different codes on the same commercial hardware demonstrate how Quantinuum's flexible error-correction architecture and highly accurate trapped-ion technology work together to enable exponential improvements in accuracy, with performance now within the target range for Apollo, our first universal fault-tolerant quantum computer planned for 2029.
Learn more: quantinuum.com/press-release…
1
111
