Jean Ibarz

Qwen 3.6 35B A3B is great. Today might be the takeoff of local models ! ❤️

@0dteezy @sudoingX It is with flash attention and KV in 4 bits

@sudoingX Aint no way this is fitting on a 3090 with that context

85-100 tok/s on the 3090 with qwen 3.6 already? that's in line with what 3.5 MoE was doing. drop your full flags and context length you tested at, i'm pulling 3.6 on the 5090 24gb and will run the same config for a direct comparison. if anyone else is running qwen 3.6 on a 3090 or any consumer card drop your tok/s, quant, and flags below. building the community benchmark sheet before i publish my own numbers

@stevibe @ivanfioravanti I tried quantized model from unsloth: didn't work at all. I than tried quantized model from abiray: works great on chat conversation and via cline through LM Studio server.

Qwen3.6 35B-A3B: smarter, but forgot how to use tools? Running 6 Bench Packs on BenchLocal across 3 open-source Qwen models. ✅ ReasonMath: 92 vs 85 vs 86 — 3.6 wins ✅ InstructFollow: 97 / 97 / 97 — tied ❌ ToolCall: 83 vs 97 vs 100 — 3.6 tanks Qwen3.5 27B still the tool-calling champ. 3.6 clearly leveled up reasoning, but tool use took a hit. DataExtract live now. BugFind + StructOutput next.

⚡ Meet Qwen3.6-35B-A3B：Now Open-Source！🚀🚀 A sparse MoE model, 35B total params, 3B active. Apache 2.0 license. 🔥 Agentic coding on par with models 10x its active size 📷 Strong multimodal perception and reasoning ability 🧠 Multimodal thinking + non-thinking modes Efficient. Powerful. Versatile. Try it now👇 Blog：qwen.ai/blog?id=qwen3.… Qwen Studio：chat.qwen.ai HuggingFace：huggingface.co/Qwen/Qwen3.6-3… ModelScope：modelscope.cn/models/Qwen/Qw… API（‘Qwen3.6-Flash’ on Model Studio）：Coming soon～ Stay tuned

On produit la même richesse, en y consacrant beaucoup moins de temps de vie humaine. Oui, les informaticiens dont je fais partie contribuent à la suppression de leur propre emploi, et c’est un bénéfice pour l’humanité. Il n’y a aucun intérêt à faire un travail qui peut être automatisé et réalisé par une machine à moindre coût. Si vous voulez payer des gens pour effectuer un travail qui ne produit rien d’utile pour la société, libre à vous de le faire, mais il n’y a aucun intérêt à les obliger à travailler.

Les informaticiens seront-ils remplacés par l'IA ? L'IA Claude répond : « Ils ont construit l'outil qui les remplace. Collectivement, les informaticiens ont construit quelque chose qui se retourne contre eux. » Voir en intégralité sur ma chaîne YouTube : youtu.be/OSAHTpRaKtw Sur une idée de Bernie Sanders.

@AlexXplore La délinquance et l'injustice galopante dans notre pays ne crée aucune richesse, mais elle crée au moins des emplois.

🇫🇷 Les actes de malveillance, incluant vols de cuivre, vandalisme et sabotages, ont causé plus de 800 000 minutes de retard et impacté près de 40 000 trains en 2024. 📡 La SNCF a une flotte de 200 drones fabriqués par la start-up française Delair pour surveiller ses 28 000 km de voies ferrées 24h/24, notamment la nuit avec des caméras infrarouges. 👀 Ces drones permettent de détecter les personnes en flagrant délit et ont déjà contribué au démantèlement de plusieurs réseaux criminels. 🛡️ L’objectif affiché est de « dissuader l’acte quel qu’il soit » grâce à une surveillance dissuasive et une intervention rapide de la police ferroviaire. 💰 Cette contre-attaque high-tech s’inscrit dans un budget annuel de 100 millions d’euros dédié à la protection du réseau ferré français. youtube.com/watch?v=4J2-fE…

Dylan Patel says GPUs are no longer the biggest bottleneck. According to @dylan522p, now CPUs are the constraint. In the early AI era, CPUs were the laggers. You used them for storage, checkpointing, pre-processing, etc. (pretty light workloads) The models weren't agentic and couldn't go step by step. Just string in and string out (simple inference) Then OpenAI launched O1 preview in September '24, and RL training loops have since tightened every month. - initially it was checking model output with regex - then running classifiers - followed by code unit tests + compilation - and finally agentic flows calling databases & scientific simulations The model outputs to an environment, gets verified, and trains on it. Coding agent revenue went from a couple billion to north of $10B in roughly 6 months. Something like Codex 5.4 can work agentically on its own for 6-7 hrs straight - doing all sorts of calls (databases, cron servers, scraping) That requires insane CPU capabilities. And over the last two quarters, the entire cloud market ran out of CPUs. - GitHub has been really unstable lately - Amazon's CPU server installations 3x'd year over year - Microsoft sold all of its spare CPUs to Anthropic & OpenAI Earlier, it was 100 megawatts of GPUs served by 1 megawatt of CPUs. Now that ratio is getting much closer for both RL training and agentic inference. There's simply no capacity anywhere, and it's causing massive instability.

@floor_per_area @AlecStapp Per article XIV, page 203 of the second amendment, structures connected via underground passage are to be counted as one building. Source: trust me, but also maybe don’t.

@AlecStapp In high demand: Duplex houses of 149sq meter each, possibly connected by a tunnel.

Government regulation in France: Above a certain size, building new homes requires a licensed architect. Outcome:

@floor_per_area @AlecStapp 🤣

Big news from Europe for Tesla!

@appealstoheaven @thsottiaux It might be ok for a freelance working alone but I can't recommend this for my company.

@_ibarz @thsottiaux That makes sense. Especially the training point. In my experience when I was a pro subscriber they had an option in settings to say you don’t want your data used for training. Is it trustworthy? I’m not sure

We did it, say hi to the $100 plan! It should be the sweet spot for a ton of you. It comes with a ton of codex usage. And yes we are resetting the limits again too as I mentioned yesterday. Let’s keep building!

@appealstoheaven @thsottiaux Sharing one account is against the rules of OpenAI. Also I think Pro plan does use data for training.

@_ibarz @thsottiaux So can you not just use the Pro plan if the business plan isn’t sufficient? Are businesses not allowed to use the plans individuals use? Set up an “IT” user at your company or something and use that to sub to the pro plan

@davicorn @thsottiaux My bad you're right

@_ibarz @thsottiaux No, it’s 20x for the $200 plan. So same as Claude

I get so many personal thank you emails for something and it’s not even a feature. All we needed apparently was a new codex plan. Can we rest now?

@davicorn @thsottiaux The quotas is proportional to the plan pricing: 1x 20€, 5x 100€, 10x 200€

@thsottiaux Would love some clarification on what multiplier the $200/mo plan is, to make a better decision :) Is it also 5x and then 20x like the competitor?

@cybersecdev_ @heeney_luke Yes they fucked us

@heeney_luke Same here , after the receipt of the “Updates to codex credit pricing “ mail, burnt through my 5hrs limit in 20mins

did codex just change their 5 hour limits? I am suddenly burning through it in an hour after never coming close before. Gah, usage limits are why I use it over Claude!

@CtrlAltDwayne @heeney_luke Business plan quotas have changed

@heeney_luke What plan are you on? I'm on the $200 Pro plan and haven't noticed any change in session limits. I've been using xhigh reasoning for the past couple of hours and have 93% remaining on my 5 hour limit.

@heeney_luke They change the rates to match API pricing, that's why. Now business plans have 4x less quotas than ChatGPT Plus plan

@Therealsteal01 @thsottiaux If you're using business plans they fucked us since rate changes since 2 April

@thsottiaux Why did you guys completely nuke my 5hr limit window? 2 tasks now consume 40% of usage??? Insane

@thsottiaux I just noticed why we hit limits in no time with business plans: you changed pricing to align with token pricing, now since 2 April business plans have much less quotas than chatgpt plus users, that's a very low blow. Not thank you👎