AMJ

#TecAddict
#Motard tôt le matin aussi
OTP Ekko Emerald #LeagueofLegends

[AMJ @Aemji_]

it's live and available #heracles https://t.co/UwCkEIksty

[Mathieu @Ma_thie_u]

Un comprimé de Paracétamol au Royaume-Uni coûte 5 fois moins cher qu'un comprimé de Doliprane en France.

Il faut le dérembourser intégralement.

[AMJ @Aemji_]

Je me rends compte qu’en Chine s’est très important
Et je me dis que cette croyance dans le fait que gagner beaucoup d’argent est corrélé à la valeur d’une personne est vrai, signifie que l’économie est globalement saine.

D’un point de vue occidental, c’est un peu différent
Je viens d’une petite île avec de FORTE disparité économique et j’ai gagné plus d’argent que la pluspart des gens.
J’ai pu ramasser énormément d’argent parce que faire mon travail était compatible avec mes valeurs, même si les autres acteurs économique s’en tapait complètement, je prennais la peine d’aider mes clients et de changer positivement leur quotidient tout en faisant mon boulot

J’ai gagné tellement de sous que j’ai pu quitté m’a petite île pour aller ailleurs, une fois arrivé à l’etranger j’ai été dans des situations où l’étique et la richesse étaient ABSOLUMENT INCOMPATIBLE, j’ai refusé et je gagne actuellement bien moins qu’à mes débuts, des « collègue » avec beaucoup moins de scrupule ont même fait littéralement fortune, mais je vis bien avec ça.
Je veux pouvoir avoir des enfants demain et savoir leur dire que je suis une personne correcte
En occident richesse et éthique ne vont pas FORCEMMENT ensemble selon les endroits, il est parfois mieux de gagner moins

[AMJ @Aemji_]

@atomic_chat_hq Can you please make a fp16 build ?

[AMJ @Aemji_]

@Youssofal_ Hi can you add offload Engram to SSD please and a 64gb compatible build ?

[AMJ @Aemji_]

@magicalchris98 @rebbie_roups Vraiment pas bête

[AMJ @Aemji_]

@danellisona I follow back every blue check, 500 Blue is the goal

[AMJ @Aemji_]

The only way to run Qwen 3.8 125b A6b 51B on 64GB, is to know wich part of the model is 51B Engram, and modify an Llm Inference Framework to instruct him to offload it to SSD
The model is designed to still work fast this way

[AppleGeek @Apple_Geek_Actu]

✨🚨 Apple annonce un jour nouveau. Bientôt.

Rendez-vous le 9 septembre à 19h pour suivre la keynote en direct.

[AMJ @Aemji_]

@ashxhart How about offload engram on ssd ?

[AMJ @Aemji_]

@UnslothAI Hello, what is the weight in GB of the Engram part ?

[Alexey Fateev @superalesha]

I've already started rewriting llama.cpp for the 51B N-grams that will be in Qwen3.8 Flash Next.

In Qwen3.8 Flash Next they are stored in a separate N-gram table, and I want to achieve one thing - offload exactly that table to RAM. The GPU will only access it for the current token.

Sounds simple? Not at all.

There are several implementation options:
1. Slow
2. Quality might get worse
3. Quantization is a whole separate problem that I won't talk about yet.

Maybe Qwen will release support for this right away. And if not - I'll do it myself.

To be continued when the weights are out.

[TheDavidTai @TheDavidTai]

Model Size Breakdown:

RAM:
- Q4 routed experts: 64.60 GiB
- Remaining BF16 model weights: 10.22 GiB
- n-gram look table: < 1 GiB
- Total: about 75.82 GiB

SSD:
- BF16 n-gram table: 95.37 GiB

[Qwen @Alibaba_Qwen]

⚡Meet Qwen3.8-Flash, a multimodal MoE and an early preview of the Qwen4 architecture, now open-weight!

The production version Qwen3.8-Flash will be available soon via QwenCloud API at just $ 0.16/1M input tokens and $ 0.47/1M output tokens.

125B parameters + 51B N-gram embeddings, with just 6B activated per token. Unmatched cost-efficiency.

What's new: 🥳
- Next architecture: GDN + QSA hybrid attention, Gated Residual, N-gram Embedding & Muon optimizer, serving as a precursor to the architecture used in Qwen4.
- Dramatically lower training and inference costs: trained at just 1/9 the cost of Qwen3.7-Plus, while outperforming it across the board with especially strong gains in coding and office tasks.
- Strong performance: scoring 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Pro, 73.9 on CoWorkBench, 84.5 on AndroidWorld, and 95.7 on MathVision (with CI).
- 262K native context, extensible to 1M with YaRN.

We’re also releasing the weights for Qwen3.8-Flash-Next, giving the community an early look at the new architecture we’re exploring for Qwen4.🚀

We can't wait to see what you build with Qwen3.8-Flash!👀👇
- Blog: https://t.co/M5hYypFLgJ
- Technical Report: https://t.co/IF0gObIkQO
- Hugging Face: https://t.co/6ow8QVAABt
- ModelScope: https://t.co/tDOn2jNuFG

[AMJ @Aemji_]

@magicalchris98 C’est mon Héros, il remet a sa place les gens complaisant d’en haut

[Jun Song @jun_song]

A lot of people who own an M3 Ultra have been DMing me asking if they should sell it.

Short answer: Yes, sell it while you still can.

I actually sold mine not too long ago.

The main reason is the new neural accelerator in the M5 series. It gives you at least 4x faster prefill speeds compared to the older M3 Ultra.

Slow prefill was always the M3 Ultra's biggest weakness. It was the exact reason why it felt slower in real-world use than a DGX Spark, even though the DGX had way lower memory bandwidth. The M5 Ultra finally fixes this.

The real-world performance gap is basically like jumping from an RTX 3090 to a 5090. If you can find a buyer right now, I strongly recommend selling.

[AMJ @Aemji_]

@ModelScope2022 It’s you Qwen-ichi 4.0 ❤️

[AMJ @Aemji_]

C’est le minimum sinon les fondamentaux c’était 0€ et faudrait shorter

[International Cyber Digest @IntCyberDigest]

❗️OpenAI has published the first performance numbers for Jalapeño, its custom inference chip that was designed with the use of AI.

On SemiAnalysis' InferenceX benchmark, OpenAI says the 700W part delivered 1.5–1.9× the throughput per watt and up to 3.6× lower end-to-end latency of Nvidia GB200 and GB300 systems.

Deployment starts inside OpenAI by year end.

[AMJ @Aemji_]

@AstraiaAI I need your help please 🥹 @luoyike2003

[Ronald Mannak @ronaldmannak]

Apple is closing the local AI performance gap with Nvidia fast and is already ahead on cost. And with Apple Upgrade leasing, it’s honestly the best option available. A few trends worth watching:

- Low end: Apple already beats systems like the DGX Spark on both price and performance. Nvidia needs a meaningful Spark upgrade to stay competitive.
- Mid range: Nvidia still leads in prefill and decode, but the gap keeps shrinking with every Apple Silicon generation.
- Total cost: Once you include the machine, power consumption, and other costs, Apple is already be cheaper than mid-range Nvidia
- Power + noise: Apple remains in a different league on efficiency and acoustics, especially important when the machine sits in the same room.
- Financing: Apple’s 0% financing/upgrade options make the economics even more compelling: use the machine for 24 or 36 months, then return it, upgrade it, or pay the remainder to keep it.
I know what I would buy

[AMJ @Aemji_]

You will regreat it

[Scott R. Leimroth @ScottLeimroth]

Apple M5 Ultra. $20,000 No thanks