Global scale infrastructure, Platforms, Data and Systems management. Engineering @RedHat; I like road trips.
London and Cambridge, England
- Tweets 18,156
- Following 1,136
- Followers 3,931
- Likes 6,499
I am going to say this clearly, the biggest impediment to Local AI community on here - is the hordes of bot accounts on twitter, using Claude and other bits of stuff to tune things they don't understand, to create hype around useless stuff and deliver token optimisation for meaningless quants and bottom rung models.
The providers literally have nothing to worry about.
1
87
Like most things in life, this isn't a blanket everyone statement. Just a highlight and the misguided algos giving a lot more credit and voice.
1
27
Karanbir Singh retweeted
Introducing SKY-001, the first Engineering World Model.
On CAD and BIM tasks, it beats GPT-6 Astra and every frontier model we tested:
-> 51-125x faster
- > 150-210x cheaper
-> Higher IOU than all frontier models
SKY-001 lets you dream a physical object before you build it.
18 months of work. Built by the team out of Maluuba's deep learning lab (acq. by @MSRResearch).
The first ‘AI General Engineer’ will be the backbone for Autonomous Manufacturing.
694
1,913
13,256
3,633,234
It's not just use of DeepSeek, it's also things made with DeepSeek and things that use DeepSeek.
If DeepSeek is a part of your SDLC, or if DeepSeek is a part of your SBOM, as an example - you can't do business with the US Govt
Wow, this is big, though not unexpected. US is removing DeepSeek from usage from all govt systems.
The 2nd/3rd order effects from this may be catastrophic though since all 3rd party systems need to be scrubbed of it and this may set a precedent/fear amongst almost every company for compliance.
And not just US companies, but any one that deals with any US company.
1
1
1
256
The Ornith models always seem to score high on benchmarks, but have never really been great at any real use for me. It feels like these guys are basically chasing benchmarks and evals on their distill
Got an early invite from @ornith_ to test their new model ahead of tomorrow's release. Really appreciate it🙏
Ornith 1.5 122B-A10B MoE!
First thing I did: ran it through the Nerfwatch Index.
20/20. A perfect score, with adaptive reasoning on.
Caveats: Nerfwatch is a narrow 20-test set built to catch regressions over time, not to rank models. This model also has a much higher reasoning setting, but a clean sweep on day one is a strong first signal.
Plenty more testing to come.
ALT nerfwatch.googz.com
1
3
278
Is this a new Muse Glimmer ?
Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware.
It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning.
Read about the model architecture, distillation recipe, quantization and more: go.meta.me/4Bcvxol
111
Karanbir Singh retweeted
Day-0 support for Ming-3.0 in vLLM-Omni
@AntLingAGI just open-sourced 5 speech models spanning TTS, audio understanding, and ASR — ranging from 0.5B to 126B total parameters (6.5B active).
Serve them today with vLLM-Omni. Congrats to the @AntLingAGI team!
Today, we're open-sourcing 5 Ming-3.0 models across 3 speech families.
- Ming-3.0-tts: 0.5B voice design & paralinguistic cues; Tiny 7.9B A1.3B long-form speech.
- Ming-3.0-audio: Flash 126B A6.5B understanding & tools; Tiny 9.3B A2.7B efficient dialogue.
- Ming-3.0-asr: Tiny 8.6B A2B dialects & context.
2
7
60
4,525
If you use @OVHcloud look at this urgently
Si vous avez téléchargé aujourd'hui la dernière version de CLI d'OVHcloud, merci de la mettre à jour à nouveau très rapidement. Sinon effacer. Il s'agit de repo:
github.com/ovh/ovhcloud-cli
La version CLI de 14h47 a été compromise par un hacker. Nous l'avons détecté et bloqué à 16h21. Merci au client qui nous a remonté le souci. La version compromise est v0.11.0 v0.12.0 v0.13.0 v0.14.0 v0.15.0.
La version v0.10.0 et la future 0.16.0 est clean.
On vient de finit les premières vérifications et aucun autre repo n'est affecté. Nous avons aussi retracé l'incident. Le hacker a utilisé le compte github d'un des développeurs et a ajouté un binaire dont le job est récupérer les secrets sur le disque. La qualité du binaire est médiocre mais il peut aller jusqu'au bout. Si vous êtes concernés, merci de changer tous les secrets sur votre disque.
On continue les recherches sur le binaire et on va publier les détails additionnels sous peu.
1
2
414
Everyone pretends to build in public, but very few people actually are. They are shimming in public and trying to squeeze some protected value.
All these open core and open whatnot's is at threat, given that LLMs can complete the needed functionality at low / no cost.
The risk and continuity conversation will mature forward too.
1
2
193
QAD's looking good
This is why NVIDIA’s QAD quants matter for local AI.
GLM-5.3-Flash-FP4-Local gets extra training to recover what compression loses. With quantization-aware distillation (QAD), the original BF16 model acts as a teacher, helping the compressed model learn to give similar answers while working with lower-precision weights.
That takes more work than a regular quant conversion: NVIDIA used 500 million training tokens over 3 epochs for this release.
The result is a 320B model using about 166 GiB, validated on two 128 GB DGX Sparks, with room for context. Quality stays close to BF16 on NVIDIA’s two reported benchmarks: GPQA 91.3% vs 92.1%; tool use 92.4% vs 91.9%.
EXL3 at 2–3 bits per weight can squeeze models into even less memory. But fitting a bigger model doesn’t automatically mean better answers. For my coding and math work, this QAD GLM build is a much better choice than the DeepSeek-V4.1-Flash 2-3 bpw EXL3 builds I tested on the same two Sparks.
I’d rather preserve more of a model’s ability than chase parameter count at the lowest possible bitrate.
More releases like this, please @NVIDIAAI !!!!
Model: huggingface.co/nvidia/GLM-5.…
How QAD works: developer.nvidia.com/blog/ho…
118
A good summary and highlights the Joe jobbing about to come next
Are you an OpenCode user?
I spent the weekend reading every 𝕏 post, reddit, and github issue looking at the impact.
So far, it appears that more users have been affected than previously reported.
Here's how to protect yourself👇
1
3
295
Karanbir Singh retweeted
Full eval results are in.
OrcaCyber Nano 1.0 scores 78.4% on SWE-bench Verified, nearly matching DeepSeek V4 Flash's reported 79.0%.
A 27B dense open-weight model designed to run locally under 16GB for cyber defense.
huggingface.co/orcarouter/Or…
🚨 Emergency release: OrcaCyber Nano 1.0 is now open-weight.
Part of our OrcaCyber model series: Nano, Zero, Ultra.
Based on our early evals, Nano is the most capable local model in its under-16GB class. 27B parameters. 262K context. Vision. Fully offline.
Early benchmarks:
AIME 2025: 96.23%
MATH-500: 98.60%
GSM8K: 97.50%
MBPP+: 84.66%
IFEval: 82.81%
Built for AI defense, vulnerability research, code security, and incident response.
We're shipping early because defenders can't wait. Cyber-specific evals coming next.
huggingface.co/orcarouter/Or…
20
24
306
24,019
Karanbir Singh retweeted
1/
The complete manuscript of my book with Weehyong Tok, LLM Customization and Fine-Tuning (@ManningBooks), is now in Early Access. The MEAP had chapters 1 to 5 since July.
Here is what the four new chapters give you.
📚 My new book, live: LLM Customization and Fine-Tuning (@ManningBooks), with @weehyong
Which adaptation technique, and how to choose? A hands-on, reproducible playbook you run end to end on one GPU.
50% off through Jul 23: hubs.la/Q04nJFBn0 #LLM #GenAI #FineTuning #AI
2
1
3
119