A bunch of LLMs scheduled to come at end of January were cancelled / delayed

Mistral Small 3 one-shotting Unsloth's Flappy Bird coding test in 1 min (vs 3hrs for DeepSeek R1 using NVME drive)

UMbreLLa: Llama3.3-70B INT4 on RTX 4070Ti Achieving up to 9.6 Tokens/s! 🚀

Claude’s reasoning model will be scary

R1+Sonnet set a new SOTA on the aider polyglot benchmark, at 14X less cost compared to o1

First 5090 LLM results, compared to 4090 and 6000 ada

The first performant open-source byte-level model without tokenization has been released. EvaByte is a 6.5B param model that also has multibyte prediction for faster inference (vs similar sized tokenized models)

I created a single-prompt benchmark (with 5-questions) that anyone can use to easily evaluate LLMs. Mistral-Next somehow vastly outperformed all others. Prompt and more details in the post.

MiniMax-Text-01 - A powerful new MoE language model with 456B total parameters (45.9 billion activated)

DDR6 RAM and a reasonable GPU should be able to run 70b models with good speed

Why I think that NVIDIA Project DIGITS will have 273 GB/s of memory bandwidth

i dont get the hype around nvidia project digits?

Llama 4 compute estimates & timeline

Are you gonna wait for Digits or get the 5090?

Now THIS is interesting

RTX 5000 series official specs

RTX 5090 rumored to have 1.8 TB/s memory bandwidth

RTX 5090 Blackwell - Official Price

Killed by LLM – I collected data on AI benchmarks we thought would last years

A new Microsoft paper lists sizes for most of the closed models

For 2025

DeepSeek does not need 5 hours to generate $1 worth of tokens. Due to batching, they can get that in about 1 minute

I don't get it.

Aider has released a new much harder code editing benchmark since their previous one was saturated. The Polyglot benchmark now tests on 6 different languages (C++, Go, Java, JavaScript, Python and Rust).

Day 10 🙂