Djamel

AI Engineer | Building production AI pipelines & local LLM systems in Switzerland 🇨🇭| Robotics enthusiast | DevOps × ML × Vision

[Djamel @BENSDJAMEL]

Tested Qwen-Image-2.0-Pro @Alibaba_Qwen with 3 prompts. Honest take:
The results are impressive.

[Djamel @BENSDJAMEL]

https://t.co/k4bU9iF8FD

[Djamel @BENSDJAMEL]

Mystery company accidentally blew $500 million on Claude AI in a single month — failed to put usage limit on licenses for employees

[Djamel @BENSDJAMEL]

5-second video generation:
Before: 25s on 8 GPUs
After: 4.2s on 1 GPU
@haoailab did that. Then open sourced it.
The bar has moved.

[Djamel @BENSDJAMEL]

Genuinely excited about SubQ — if O(n) scaling at frontier delivers, it changes everything for long-context. That said, 3 benchmarks isn’t enough. Need MMLU-Pro, GPQA, BBH. what’s the story behind that 83% research vs 65.9% prod gap on MRCR? optimistic, want API access to test

[Alexander Whedon @alex_whedon]

Introducing SubQ - a major breakthrough in LLM intelligence.

It is the first model built on a fully sub-quadratic sparse-attention architecture (SSA),

And the first frontier model with a 12 million token context window which is:

- 52x faster than FlashAttention at 1MM tokens
- Less than 5% the cost of Opus

Transformer-based LLMs waste compute by processing every possible relationship between words (standard attention).

Only a small fraction actually matter.

@subquadratic finds and focuses only on the ones that do.

That's nearly 1,000x less compute and a new way for LLMs to scale.

[Djamel @BENSDJAMEL]

@alex_whedon This is exciting — O(n) scaling at 12M tokens could be great. Would love to see MMLU-Pro/GPQA/BBH scores though. The 17pt research→production gap on MRCR is interesting. want to test it on real tasks. Anyone in early access?

[Djamel @BENSDJAMEL]

@DanielSmidstrup Company owners, we pay for it.

[Djamel @BENSDJAMEL]

@sama Makes sense - smartness unlocks new capabilities entirely, while speed/cost just make existing ones more accessible. Though at some threshold, 10x cheaper might matter more than 10% smarter?

[Djamel @BENSDJAMEL]

@MrBeast Like

[Djamel @BENSDJAMEL]

@MrBeast 1 like

[Djamel @BENSDJAMEL]

@paddix @digitalocean Having Near Opus 4.6 coding performance on your own inference engine is amazing. Well done.

[Djamel @BENSDJAMEL]

@dsikka84 @RedHat_AI 0.33% accuracy drop from W4A16 with 99.6% recovery is exactly what production quantization should look like. Excellent work.

[Djamel @BENSDJAMEL]

@paddix @digitalocean @ArtificialAnlys Amazing work on this! The NVFP4 quantization approach is fascinating. I’d love to learn more about how the accuracy holds up in production workloads compared to the benchmark evals.

[Djamel @BENSDJAMEL]

The Gemma $200,000 Challenge deadline is coming up. Anyone here building something they want to share? Always interesting to see what creative applications people come up with

[Google for Developers @googledevs]

Build for impact. Win from a $200,000 prize pool. 🏆✨

Join the Gemma 4 Good Challenge to create solutions for health, education, global resilience, digital equity, and AI safety. With a $200,000 prize pool and multiple technical tracks, discover how to scale impact using Gemma 4.

Submit your project by May 18 → https://t.co/uTgCcTNIGh

[Djamel @BENSDJAMEL]

@felangelov Curious about this take - is it the verification burden, or does AI code tend to miss architectural considerations you’d catch while writing?

[Djamel @BENSDJAMEL]

@RedHat_AI @mgoin_ @vllm_project CUDA 13.0 + PyTorch 2.11 migration in production will be the real test. Excited to hear deployment strategies and how teams are handling the dependency updates

[Djamel @BENSDJAMEL]

@bnjmn_marie Maybe adding:
•Qwen/Qwen3.6-27B-GPTQ-Int4 - widely used, good vLLM support baseline

[Djamel @BENSDJAMEL]

NVIDIA’s new Nemotron multimodal model: 256K context window (one of the largest available). English-only for now, multilingual coming soon.
Already supported on day 0 by @UnslothAI and @vllm_project 🚀

[NVIDIA AI @NVIDIAAI]

Meet Nemotron 3 Nano Omni 👋

Our latest addition to the Nemotron family is the highest efficiency, open multimodal model with leading accuracy.

30B parameters. 256K context length. 🧵👇

[Djamel @BENSDJAMEL]

@MansteinGeorg @karpathy How was his impression after the drive?

[Djamel @BENSDJAMEL]

@cinamarina I’d love to hear your 400k story — it could inspire a lot of folks here who are just starting out.

[Piotr Żelasko @PiotrZelasko]

Today we released Nemotron-3-Nano-Omni-30B-A3B - our first Omni model, with speech and audio understanding capabilities powered by parakeet-tdt-0.6b-v2 encoder.
🫡1st position on VoiceBench
🌏English only
🎙️5.95% WER on Open ASR Leaderboard
📽️Video+audio understanding