Murali Kondragunta

AI something @Aleph__Alpha. Post training LLMs

[Murali Kondragunta @gitlostmurali]

@IFM_AI Cool stuff!

On a similar note, we scaled to 25k concurrent sandboxes and ~5M rollouts. Ran into many of the same challenges along the way 😀

[Aleph Alpha @Aleph__Alpha]

Scaling RL environments and their infrastructure comes with several challenges. We documented our experiences while building Kolibri: stalling GPUs, images filling the disks, crashed sandboxes scored as wrong answers, and reward hacks.

[Murali Kondragunta @gitlostmurali]

@GenAI_is_real @IFM_AI On a similar note, we scaled to 25k concurrent sandboxes and ~5M rollouts. Ran into quite a few similar challenges and learned some lessons along the way 😀

[Aleph Alpha @Aleph__Alpha]

Scaling RL environments and their infrastructure comes with several challenges. We documented our experiences while building Kolibri: stalling GPUs, images filling the disks, crashed sandboxes scored as wrong answers, and reward hacks.

[Murali Kondragunta @gitlostmurali]

Blog link: https://t.co/dqIrmP9Pi8

[Murali Kondragunta @gitlostmurali]

Letting sandbox creation choke GPU utilization is one of the costliest mistakes in async RL

We learned this the hard way during Kolibri's development. Tackling this and several other scaling challenges helped us reach 25k concurrent sandboxes & millions of sandboxes overall

[Aleph Alpha @Aleph__Alpha]

Scaling RL environments and their infrastructure comes with several challenges. We documented our experiences while building Kolibri: stalling GPUs, images filling the disks, crashed sandboxes scored as wrong answers, and reward hacks.

[Murali Kondragunta @gitlostmurali]

We also share some cute examples of Kolibri reward hacking along the way

[Murali Kondragunta @gitlostmurali]

The move from external services to self-hosting didn't stop at sandboxes. We followed a similar path with other environment infrastructure, adding more capabilities along the way.

You can see how the stack evolved

[Murali Kondragunta @gitlostmurali]

We started with external sandbox providers but quickly moved to self-hosting our own sandbox service

This gave us much more control over the infrastructure and better observability into what was happening under the hood

[Aleph Alpha @Aleph__Alpha]

Every architectural choice has a cost. Choose where to pay it.

@alessio__serra and @PitNeitemeier derive from first principles how design choices allocate parameters, FLOPs and KV cache, compare open models, and share some of the tooling we used to design Kolibri.

[Murali Kondragunta @gitlostmurali]

During its development, Kolibri surprised us in ways we didn't expect with reward hacking & escaping sandboxes. We were a little proud but also increased our sandbox isolation

For an 80B-A3B model, it is unexpected. All the details are in "Lessons learned" section in tech report

[Aleph Alpha @Aleph__Alpha]

Small bird, fast wings, Kolibri is here.

78B parameters. 3.46B active. Up to 1M tokens of context. Built in Europe.

Now the weights are yours. Run it on your own hardware, under Apache 2.0.

[Murali Kondragunta @gitlostmurali]

This is so amazing to see! Love open source for this!

Btw, we didn't train Kolibri on blender or artifact generation. So good to see that model generalized to do this

Looking forward to more such experiments :) :
https://t.co/bS5oF6CkwT

[Fahd Mirza @fahdmirza]

💥 Kolibri is HERE, and it's Europe's answer to open-weight AI

🐦 Aleph Alpha's new German/English model: 78B total parameters, only ~3.5B active per token

🔹Up to 1M token context window
🔹Apache 2.0 license, run it on 2x H100 or 1x H200
🔹Built on a "training as code" pipeline: the big run is just the last step after hundreds of small experiments
🔹Strong on math and German-language tasks (benchmarks are the company's own claims)
🔹Tested hands-on: a full Blender + Godot trampoline park game, a cinematic doner kebab in pure HTML, and a brutal Hegel German-English philosophy challenge

🔥Watch the full video below:👇

https://t.co/SyqE5Jo2Xq

@Aleph__Alpha

[Marco Chiappetta @elchiapp]

People saying "it's worse than Qwen3.6 35B" don't understand that support comes from the bottom. If we don't support the local AI ecosystem, no matter the stage, no matter the location, we're always going to fall prey to "yeah but America/China does it better". Those who tried to build something the day GPT tool calling came out (🖐) know very well the pain of getting malformed tool calls and the model going haywire with the most trivial requests. That was just 3 years ago. Today you can run an even better model on a laptop.

I'm a huge fan of frontier labs, they are the pioneers of what we have today and I'll forever be grateful, but local AI is one of the most important (and frankly exciting) challenges of this decade. We're all in it together. Supporting each other is the least we can do. So congrats on the release! 🇪🇺

[Murali Kondragunta @gitlostmurali]

Thanks Max! ❤️

Max played an instrumental role in building the core training teams before leaving to found Construct Lab

Try out Kolibri at
https://t.co/fP920qAcT0

[Max Meuer @meuer_max]

Big congrats to my friends at Aleph. Glad I was still around when this team formed and got to see how much work went into building it end to end.

https://t.co/VVpZLOM4rK
We will offer inference for free for the next ~48 hours.

Hit me up if you are interested in bringing this into production or post-training it.

[Max Meuer @meuer_max]

Big congrats to my friends at Aleph. Glad I was still around when this team formed and got to see how much work went into building it end to end.

https://t.co/VVpZLOM4rK
We will offer inference for free for the next ~48 hours.

Hit me up if you are interested in bringing this into production or post-training it.

[Murali Kondragunta @gitlostmurali]

@bjoern_pl And because of the group size, we will have to generate 8 rollouts for each sample. And also 2 sandboxes per rollout (for some envs) as we do grading in a fresh sandbox

[Murali Kondragunta @gitlostmurali]

Our model, Kolibri, is out! ���

Incredibly proud of how much we've scaled in just a few months

RL is as much an infrastructure challenge as a modeling one. We've scaled our RL environments and their infra to support 20K+ concurrent sandboxes per training run and 100K+ sandboxes overall. More details in the report

A massive engineering effort behind the scenes. Proud of what we've built!

[Aleph Alpha @Aleph__Alpha]

Small bird, fast wings, Kolibri is here.

78B parameters. 3.46B active. Up to 1M tokens of context. Built in Europe.

Now the weights are yours. Run it on your own hardware, under Apache 2.0.

[Murali Kondragunta @gitlostmurali]

@bjoern_pl Sandboxes are for SWE environments - which have a lot of samples

[Murali Kondragunta @gitlostmurali]

@Aleph__Alpha It's a good model ser 🫡