Samir

@scaling01 🦞²

the race is on

Fable 5.1 next week GPT-6 within the next 1.5 months

@quxiaoyin Remember AI 2027? We are so much faster

Kimi just surpasses Fable across multiple benchmarks, and it’s open weight! China has caught up.

@scaling01 But 3/15

Kimi K3 scores 1687 on GDPval-AA v2 higher than Opus 4.8, still behind GPT-5.6-Sol's score of 1,747.8

@scaling01 @eliebakouch Never underestimate chinese dev's

@eliebakouch it means someone could easily post-train on top of it and build open Mythos

IT'S HAPPENING any second now

@kimmonismus With GPT Work they created an user experience similar to just "talking with ChatGPT". Not just 5.6 beeing a good model but bringing agentic capabilities to an non-tech user friendly level.

This is crazy. The pressure from OpenAI seems to be really intense. It's truly rare for Anthropic to reset every 5 hours *and* weekly. The only likely reasons for this are the success of the Codex, its growth, and the repeated reset to version 5.6.

@signulll The internet becomes an agent place. No need for humans.

we’re entering the end stage of the address bar, domains, tabs, & all the stupid dumb tech plumbing browsers exposed to normal users. i.e. i almost never begin at the browser layer anymore. the journey starts in claude, chatgpt, siri, or some other ai system. which means the browser has been demoted from destination to runtime. you still might land there via links, checkout, dashboards, authentication, & documents. but you rarely *start* there. websites are no longer competing primarily to be visited. they’re competing to be retrieved, cited, acted on, & represented correctly by agents. traffic becomes an implementation detail. which is why killing atlas was probably one of openai’s smarter decisions & i have written about it ton a year + ago (search my tweets). the browser itself will matter less & less. the systems that understand intent, choose where to go, & act on your behalf will matter far more.

@effi288 Great Work, there is so much work to do in europe when it comes to ai-tech. And independence will be necessary.

My favorite plot of the whole project: watching SOOFI-S climb past the other models during training, checkpoint by checkpoint. Max’s thread on how we evaluated — 1M+ benchmark runs, all open 👇

📄 Releasing the Soofi S pretraining tech report: a sovereign, open foundation model for German and English Today we’re publishing the full pretraining tech report and project page for Soofi S 30B-A3B — a Mixture-of-Experts hybrid Mamba model trained on ~27 trillion tokens with deliberately up-weighted German. What’s in the report: 🏆 Strongest fully open model in our evaluations on BOTH the English and German aggregates — ahead of Olmo 3 32B and Apertus 70B (full methodology in the report) 📋 Radical transparency: complete per-source data accounting, all hyperparameters, training + eval code, checkpoints — everything under permissive licenses 🇩🇪 Trained end-to-end on Deutsche Telekom’s Industrial AI Cloud in Munich — sovereign AI infrastructure on German soil Soofi S combines frontier-level capability with the highest measured aggregate long-context decode TPS, and unlike full-attention dense baselines maintains high throughput as context grows. The figure plots Capability Index versus measured aggregate decode TPS/GPU at 40K context and batch 32. The Capability Index averages five benchmark groups, i.e., Code, GSM8K, GPQA-Diamond, English aggregate, and German aggregate, after normalizing each group to the best plotted model. Aggregate decode TPS/GPU is measured with a TP=1, one-B200 vLLM latency-subtraction protocol.

@intheworldofai And they stopped distilling from US-Models. They start to come close enough to US-Models so they can price dump. Atleast Anthropic will run into compute problems also by gaining more and more users. Coming Bottlenecks: Compute -> Humans Able to use AI -> Energy

Tencent’s cooking HY4 right now and it’s stacking way more params than the last one (that 295B/21B beast). Multimodal too. They’re dead set on cracking the top tier of Chinese models by 2027. Meanwhile ByteDance Seed is cooking an absolute monster model of their own. China AI arms race just got spicy as hell

@quxiaoyin It might work short/mid term, if they can out accelerate open-source models in this phase. Anyways open source is the future - long term.

turns out USA is adopting a different strategy than open-source. Make models closed-source but cheaper

@goodside That's benchmarking

Claude Fable 5 (Max) creates a crossword puzzle of the first 150 Pokémon:

@alexandr_wang @scaling01 He didn't see that one coming

@scaling01 we’re working on it

is Meta regarded? just put Muse Spark 1.1 on openrouter you dipshits

@kimmonismus Okay so we have Palantir + Legal Mass Surveillance + growing dependency on US What could go wrong?

To all my German followers: This is how the new chat control system works. The EU has introduced mass surveillance.

@DanDr1s So the new model appeared in Arena yesterday is Opus5 most likely?

🚨 Claude Opus 5 will be releasing somewhere this month as a response to GPT 5.6 sol, and grok 4.5 . Don’t believe me? - I was the first to call out the release date of Claude Sonnet 5

@reach_vb Super-Apps + Bidi + 750 tokens/s = Generative UI?

LFG!!!

@kimmonismus You're either (most likely) the cheapest or the best. He won't achieve either-Chinese models will likely be both cheaper and smarter. But with half of humanity as his customer base, most of them probably won't need advanced reasoning anyway.

Last week, it was still a race between Anthropic and OpenAI. One week later, it is a race between Anthropic, OpenAI, SpaceX, and Meta.

Meta is releasing an agentic and coding model that can compete with GPT-5.5 and Opus 4.8? And on top of that, it is SOTA on HLE? Holy, that was not on my bingo card. Alexandr Wang cooked. What is happening today!

@scaling01 You're either the cheapest or the best. He won't achieve either-Chinese models will likely be both cheaper and smarter. But with half of humanity as his customer base, most of them probably won't need advanced reasoning anyway.

the frontier can basically charge any price it wants

Zuck is ready to throw hands he wants to pull Anthropic and OpenAI into a price war Meta has enough revenue and profits OpenAI and Anthropic don't

@kimmonismus 40% confidence, 40% image, 20% F U

Hold on: Is China now doing the same thing as Anthropic, just the other way around, by banning U.S. companies from accessing Chinese frontier models? China’s confidence seems to be growing by the day. Apparently, they believe they no longer need distillation on this scale to develop frontier intelligence. Instead, they seem more concerned that the U.S. could now steal their intelligence.

@GoogleDeepMind What a time to be alive

@PaulGugAI @kimmonismus Just if closed-source does not out-accelerate them next month's. Anyways Open-Source is the long term winner for general use worldwide.

@kimmonismus Perhaps the gap between frontier and equivalent open source will also narrow over time also?

Frontier labs have shifted into rolling-release mode since mid last year. In 2023 and 2024, flagship launches still felt spaced out: GPT-4, GPT-4o, Claude 3, Claude 3.5, Gemini 1.5. From 2025 onward, the cadence tightens sharply. OpenAI moves through GPT-5 and the 5.x line. Anthropic stacks Opus 4, 4.1, 4.5, 4.6, 4.7, 4.8 and Fable 5. Not only are the models getting faster and better, but the speed of releases is also increasing.

@ant_dracula @petergostev clawbuis.com

@SamirBallhausen @petergostev No you don't.

We are now in a position where a tiny proportion of the population uses Fable or soon GPT-5.6, while everyone else's experience of AI is 8-30b-model level - Google's AI Overviews, Meta AI, ChatGPT free tier, maybe MS Copilot at best. People outside of tech must be completely baffled how this is supposed to take their job, and annoyed that hundreds of billions are being poured into it.