Re-derives LLM scaling laws, tracing prior disagreements to how compute budget was modeled, then trains 7B and 67B models on 2T tokens. The 67B model beats LLaMA-2 70B on code, math, and reasoning; its chat variant tops GPT-3.5 on open-ended evals.
Reworks the MoE layer to push each expert toward a narrow specialty: split experts into many finer ones and activate more per token, plus reserve a few always-on shared experts for common knowledge. A 2B model matches GShard 2.9B; at 16B it rivals LLaMA2 7B on ~40% of the compute.
A family of open code models (1.3B-33B) trained from scratch on 2T tokens of project-level code, using a 16K-window fill-in-the-blank objective. Beats Codex and GPT-3.5 on code benchmarks and ships under a license permitting commercial use.
Reaches 51.7% on the competition-level MATH benchmark with a 7B model and no tools or voting, rivaling Gemini-Ultra and GPT-4. Built on a 120B-token math corpus mined from Common Crawl, and introduces GRPO, a memory-efficient PPO variant for reasoning.
A 236B-parameter Mixture-of-Experts LLM that activates only 21B parameters per token, cutting training cost 42.5% versus a dense 67B model and shrinking the KV cache 93.3% via Multi-head Latent Attention, with 128K context.
Open-weight Mixture-of-Experts LLM with 671B total parameters but 37B activated per token, trained on 14.8T tokens for 2.788M H800 GPU-hours. Matches leading closed models at a fraction of typical training cost via FP8 and architectural tricks.
Shows that LLM reasoning can be incentivized through pure reinforcement learning, with no human-annotated reasoning traces. Self-reflection, verification, and strategy-switching emerge on their own, and the patterns transfer to distill smaller models.
Compares standard human psychometric questionnaires (PVQ, BFI) with generation‑based profiling to test whether questionnaires predict real LLM responses. Finds big divergences: questionnaires exploit lexical cues and elicit alignment‑consistent answers, mischaracterizing LLM behavior on everyday queries.
An open large language model pairing DeepSeek Sparse Attention (DSA) for cheaper long-context inference with a scaled RL pipeline. Authors claim parity with GPT-5, with a high-compute Speciale variant surpassing it and rivaling Gemini-3.0-Pro on reasoning.
Provides a conditional memory module that performs O(1) N‑gram lookups and fuses static embeddings into transformer hidden states — enables offloading large embedding tables to host memory with minimal inference overhead.
Preview of an MoE model family (V4-Pro: 1.6T params, 49B active; V4-Flash: 284B, 13B active) built for 1M-token contexts. A hybrid attention design cuts single-token inference FLOPs to 27% and KV cache to 10% versus V3.2 at million-token length.
Introduces Draft-OPD, an on-policy distillation method for training lightweight draft models used in speculative decoding — it focuses learning on draft-induced errors via target-assisted rollouts and replay, improving acceptance length and enabling >5× lossless LLM inference acceleration.