Tag
Open Weight Models on The Latent Daily: 3 stories and 18 models. Latest: Budget the tokens: how finance teams are starting to plan enterprise AI spending.
Dell's CFO says finance chiefs now give token spend its own budget line. Here is how companies can track, forecast and cap AI costs without slowing teams down.
thelatentdaily.com Editorial Team4 min read
What matters when buying a computer to run language models locally: memory, GPU, Apple silicon, quantization and model size, with setups for each budget.
thelatentdaily.com Editorial Team4 min read
Mistral AI, founded in Paris in 2023 by ex-DeepMind and Meta researchers: open-weight models like Mistral 7B and Mixtral, Le Chat and its European strategy.
thelatentdaily.com Editorial Team3 min read
A 27B dense vision-language model from Alibaba's Qwen3.8 generation, aimed at coding, professional work and long-horizon agentic tasks.
Google DeepMind's fourth generation of open models, the first under Apache 2.0, with dense and mixture-of-experts sizes, configurable thinking and native multimodal input.
Mistral AI's 119B mixture-of-experts model that merges the Instruct, Magistral reasoning and Devstral coding lines into one model with switchable reasoning.
NVIDIA's open 30B hybrid Mamba-Transformer mixture-of-experts model with about 3.5B active parameters, built for efficient reasoning and agents.
Ai2's fully open 7B and 32B models with Instruct and Think variants, released with their training data, code, checkpoints and recipe.
Baidu's first open-weight ERNIE release: ten models from 0.3B dense to 424B mixture-of-experts, including vision-language versions.
Alibaba's Qwen3 family of dense and mixture-of-experts models from 0.6B to 235B parameters, with a switchable thinking mode for reasoning.
IBM's enterprise-focused open models with 128K context, reasoning support and Apache 2.0 licensing, trained on data IBM documents for business use.
AI21 Labs' hybrid Transformer-Mamba mixture-of-experts model with 256K context, efficient for long-document RAG.
The Technology Innovation Institute's family of small open models (1B to 10B) designed to run on laptops and light infrastructure.
LG AI Research's bilingual English-Korean models in 2.4B, 7.8B and 32B sizes with 32K context.
The Allen Institute for AI's fully open language models: weights, training data, code and intermediate checkpoints are all released.
AI2's fully documented post-training recipe applied to Llama 3.1, with open data, code and evaluation for every stage.
Cohere For AI's multilingual research models covering 23 languages, built to close the quality gap between English and other languages.
Alibaba's Qwen2.5 family of dense models from 0.5B to 72B parameters, strong in coding, math and multilingual tasks, widely used for fine-tuning.
Nous Research's fine-tunes of Llama 3.1, focused on role-play, function calling, structured output and following the system prompt closely.
A 12B model built by Mistral AI with NVIDIA, with 128K context, strong multilingual ability and a new efficient tokenizer.
Mistral AI's sparse mixture-of-experts model: eight experts, two active per token, matching much larger dense models at launch with lower inference cost.