🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

⚡️ Reka Flash ⚡️, our 21B multimodal assistant, is much better now 🔥🔥🔥 Try it for free at chat.reka.ai. No need to sign up anymore!

Dani Yogatama, our CEO, had the opportunity to share about the work we do at Reka, our multimodal AI models, his perspectives of the future of AI, as well as how we use AMD Instinct MI300X and ROCm software at AMD's Advancing AI event.

I wish there was a more direct question about keeping ARR vs. getting rid of it vs. making it optional, but it's still great that we have a chance to give our opinion. Please don't forget to vote!

Excited to finally launch Space in private beta! Join the waitlist at reka.ai/space

📢 Make multiple choice benchmarks more difficult by replacing a random choice with "None of the above"! Interesting to see how certain models are a lot more affected (cough cough Qwen 7B)... What does a stellar MMLU mean when such a trivial change causes a drop of 20 points? 🤔

Looking for a challenging benchmark? BIG-Bench Extra Hard comprises 23 of them... and our linguistic reasoning benchmark, Linguini, is the hardest of them all, with all models scoring below 20!

Open sourcing our 21B reasoning model! Trained from scratch, strong numbers, Apache 2.0... Go build with it! 🔥🔥🔥 And even more excited to announce Nexus, our new platform to create AI workers, available in private preview. More coming soon!

Reka Flash 3 update: added another great model Gemma-3 27B (instruction tuned version) to our benchmark results since many people asked us how they compare on these datasets. Also Apache 2.0 vs Gemma license.

Reka Flash 3 is the best open source model on BALROG!

After multiple requests, here it finally is: Reka Flash 3 is an amazing model on my vibe eval! While its consistency is pretty bad, the pass@5 is crazy, on par with GFT in general and on par with Claude 3.5 in coding. And thats 21B w/ Apache 2.0 license!

Reka Flash 3 is the highest scoring model of its size according to Artificial Analysis! 🔥🔥🔥

Euskarazko ChatGPT ireki bat nahi? HiTZ zentroa (UPV/EHU)-n horretan ari gara eta azken bultzada emateko zure laguntza behar dugu! Oso erraza da parte hartzea: 1. 🌐Sartu ebaluatoia.hitz.eus helbidean 2. ❓Egin zure galdera 3. 🩷Aukeratu 2 erantzunen artean onena 4. 🎁 Irabazi sariak

Another account of Reka Flash 3 doing strongly in the wild! 💪 "Reka Flash 3 is really capable, especially for its size. In our testing, it came close to the performance of 32B models, such as the R1-distill of Qwen and QwQ 32B"

Vision capabilities are core to our mission to develop multimodal AI that understands both the digital and physical worlds. In our latest blogpost, we discuss our approach to process video inputs and our industry-leading results from different domains. reka.ai/news/reka-visu…

Including English data when adapting LLMs to new languages doesn't affect perplexity, but can have a huge impact in downstream tasks 🤯 What causes this unusual uncorrelation? Why is English data so important? 🤔 Check out our new paper where we solve the mystery! 🧵👇