TNG Technology Consulting GmbH

Today we release DeepSeek-TNG R1T2 Chimera. This new Chimera is a Tri-Mind Assembly-of-Experts model with three parents, namely R1-0528, R1 and V3-0324. R1T2 operates at a sweet spot in intelligence vs. output token length. It appears to be... * about 20% faster than R1, and more than twice as fast as R1-0528 * significantly more intelligent than R1 in benchmarks such as GPQA Diamond and AIME-24/25, albeit not quite on R1-0528 level * much more intelligent than our first R1T Chimera, and also think-token consistent, which is a major improvement We perceive it as generally well-behaved and a nice persona to talk to. The weights are on @huggingface under the MIT licence. We are looking forward to your experiments and feedback! Thanks to @deepseek_ai for giving their models to the world, to @chutes_ai and @openrouter for hosting R1T, to @WolframRvnwlf for benchmarking it, to @xlr8harder for beta-testing the new Chimera, and to @natolambert for constructive discussions at @aiDotEngineer.

@natolambert If you like, you can spend some months here in Munich and we build something meaningful.

Very happy for Karpathy. Also very lonely times in open science.

Thanks for doing interesting and intriguing research! "Wir müssen wissen, wir werden wissen" (David Hilbert)

just want to do a quick shoutout for all supporters: Big thanks to: @PrimeIntellect @tngtech @hillclimbai @runpod and @0xSero who made the original shoutout and is also supporting. ^all very cool people/companies :) dm me if you want to support my research :) (more compute is always welcome lol)

Our robots are in the metal hospital right now.

@Yuchenj_UW Real-time measurement:

GPU shortage is worse than ever. H100s cost more today than they did 3 years ago, and you cannot get them on-demand. The big AI labs have locked up most of the supply for years. I’m worried university researchers and individual developers simply won’t be able to get GPUs.

@0xSero What kind of compute? 4x/8x B200? Would 8xRTX6k Pro work?

I believe this guy is onto something really interesting and valuable. I will commit 500$ for compute if I can get any of the clouds match me (Lambda, Prime Intellect, Hotaisle) He’s working on continuous learning and has given lots of valuable advice. 🙏

@nicolaygerold @xeophon Sent us a DM.

@tngtech @xeophon Who do I have to write / where to signup? :)

I will give a talk at TNG‘s Big Techday about evaluations next week! Would love to meet some of y'all there :)

@xeophon @nicolaygerold Ticket AND Coffee :-)

@nicolaygerold Yes, let’s do it! I am sure @tngtech has a ticket for you if you want one :)

@allen_ai What a beautiful idea! Congratulations!

Today we’re releasing EMO, a new mixture-of-experts (MoE) model trained so modular structure emerges directly from data without human-defined priors. EMO can use a small subset of its experts for a given task while keeping near full-model performance. 🧵

@tugot17 One of the big ones. Whether to believe them is a difficult choice.

@tngtech who sells NVL72 with delivery in July 😅?

GPU acquisition prioritization, your opinion:

@tugot17 GB300, yes, supposedly fast delivery until July 1st - but is that true? VR200, yes, supposedly January 27 delivery - if VR does not have the amount of teething troubles that GB200 had

@tngtech Can you actually buy one?

New walking policy for our walker.

@justanotherlaw @ben_sturgeon Thank you for your very interesting analysis!

A recent viral paper claims to reverse-engineer the parameter counts of frontier models: GPT-5.5 = 9.7T, Opus 4.7 = 4.0T, o1 = 3.5T, etc. @ben_sturgeon and I investigated and found serious issues in the paper; fixing them gives GPT-5.5 as ~1.5T (90% CI: 256B-8.3T).

If you know where to buy B300 servers for $0.5 million "in the West", give us a call. The street prices that we get quoted in Europe are already in the range of $0.8-1.0M, both air- and watercooled, without the rack and the Infiniband switch and the CDU. There seems to be very strong B300 demand in the US, the EU et al, too.

The price of B300 servers in China has risen to the point that it is now double the price in the West. The crackdowns on smuggling appear to be having an effect. Mythos has changed things.

@tngtech @Zai_org @opencode 0xsero.github.io/blackwell-gpu-… <—— I address all this (lots of slop but the data comes from months of digging)

Experiments using @Zai_org 's GLM-5.1 on Hopper and Blackwell, using @opencode with 50k input tokens and 500 output tokens on average. The RTX 6000 Pro x 8 node seems not to scale well with concurrency, while the B200 x 8 is just very strong. Are these measurement errors or reality? With luck and new kernels, how much could the RTX be improved?

@latkins @eliebakouch FTW - For Trinity's Win :-)

@eliebakouch Do me a solid bro and add another model to that comparison

new mistral model: 128B dense with an arch from 3 years ago (llama 2), very low context (128k), priced higher than deepseek v4 pro (1.6T total params, 1M context) and every other oss model that outperforms it this is very sad

@0xSero So what are the chances of finding a workaround? Sometimes a cliff can be scaled.

Nvidia, why?

Registration for #MuniHac 2026 is open! #registration" target="_blank" rel="nofollow noopener">munihac.de/2026.html#regi… #MuniHac will take place October 9–11, 2026 at the @tngtech office in Munich. Keynote by @TacticalGrace confirmed, more to come. Want to contribute? Check out our CfC #cfc" target="_blank" rel="nofollow noopener">munihac.de/2026.html#cfc! See you in Munich!