Meta

In the team's own measurements, 4 MLPerf Tiny models each ran in 21.5ms or less and under 90KiB of RAM on a single Cortex-M55 core of the Alif Ensemble E8 (the anomaly detection autoencoder in 5.2KiB). The team noted the

Meta claims it is the most capable US open-weights model to date. It adds safety post-training for standalone use on top of the same base model as Muse Spark 1.3, and ships with SGLang and vLLM optimizations and Helion k

MSL keeps to its closed-API route while putting agent efficiency (fewer tool calls and tokens) forward as its headline metric

The first open weights since Llama 4, released under Apache 2.0; an on-device agent at under 20GB in 4-bit

The point at which Meta's two-track strategy was settled: the frontier model closed, with only separate small models open

The point at which the single-controller model proposed by Google's Pathways in 2022 entered the core PyTorch lineage. TorchTitan becomes an actor inside Monarch, and torchforge is built on top of Monarch

Just as Triton abstracted CUDA by one level, Helion abstracts Triton by one more. The next layer in the lineage of kernel authoring

A shift from evaluating static information retrieval to measuring agent robustness in dynamic, asynchronous environments

A turning point for the Llama line; Behemoth's release postponed

The mismatch between versions submitted to leaderboards and the publicly released ones became a public issue

A turning point for trust in the Llama brand

Processes bytes directly, without a fixed vocabulary, using entropy-based dynamic patches; the first large-scale scaling of the tokenizer-free approach

An update to the Llama 3 line that shipped a single 70B instruction-tuned model, with no separate pretrained release

The point at which a PyTorch-native training framework became established as an alternative to Megatron-LM. FSDP2, torch.compile and FP8 were bundled into a single stack

The 11B and 90B were the first Llama models to support image input, released alongside 1B and 3B models aimed at on-device use

The first primary source to publish quantitative data on fault tolerance in large-scale training. Settled, for the whole industry, that checkpointing and automatic recovery are essential

Frontier-class weights were released for the first time, and the license allowed training on the model's outputs

A case showing that kernels must be rewritten for a specific generation of silicon. Quantitatively rebutted accuracy concerns about FP8 attention

The 8B delivered the previous generation's 70B-class performance, raising the bar for practical local use

On 466 questions, humans scored 92% versus 15% for GPT-4 with plugins; shifted the focus of evaluation from hard exams to practical ability

Structured self-checking into 4 steps to reduce hallucinations

A code-specialized line trained on 500B tokens of code and code-related data on top of Llama 2; a 70B version was added on January 29, 2024

Popularized open weights, and the starting point of the debate over whether it is open source

The turning point at which large-scale training became possible with the standard framework, without DeepSpeed

The first large attempt to distribute frontier-class weights after vetting

Learns on its own, through self-supervised learning, when to call APIs and with what arguments — the route of building tool use into the weights

A redesign of FSDP1. The precondition that made FP8 training, partial freezing and compiler optimization possible