Tencent chief AI scientist Shunyu Yao and the Hunyuan team held a 120-minute Reddit AMA addressing questions about model scale. The team argued that well-built smaller models like Hy3 and Hy4 can outperform larger ones at lower cost, but acknowledged Hy4 has substantial room for improvement through post-training and large-scale reinforcement learning, and confirmed plans to train larger-scale models while considering 24B-31B dense variants.
Dense Model
稠密模型
Architecture activating all parameters per inference, contrasted with sparse MoE; the team is considering 24B-31B variants.
Curated Intelligence
Google's ASA and Ranking AI Research teams optimized the open-source TPU stack (vLLM, Tunix, MaxText, Pallas) for RL training of a 0.6B dense model on TPU v7, cutting the full rollout-to-update cycle from roughly 100 minutes to about 56 seconds (~183K tokens/s), a 107x speedup while retaining the full 8,192-token response budget. All changes were upstreamed, so developers can reproduce the path via the merged tpu-recipes recipe without redoing bring-up work.
Qualcomm argues that on-device AI is moving from single-purpose features such as imaging and speech transcription toward agentic experiences capable of intent understanding and task orchestration. Its approach pairs an 'on-device first' cloud-edge continuum with Mixture-of-Experts sparsity to cut per-inference compute, backed by NPU hardware and a full-stack software ecosystem.
Nous Research has open-sourced Arete-27B, a Qwen3.8-27B fine-tune trained with plain supervised fine-tuning on 52.8M tokens using Inkling as teacher, with no RL or architecture changes. On its self-built 803-prompt eval, propaganda-style answers dropped from 22% to 0.9%, while factuality mean rose only from 2.32 to 2.98 out of 5 and still trails Muse Glimmer 30B; 216 of 803 answers still score 0, and both eval and training data are self-constructed and self-judged.
Ricursive Intelligence, founded by former Google researchers, released Ripple-65B Preview, a dense open-weight model pretrained on 12T tokens with 1M context that it claims outperforms the much larger Meta Muse Spark Open 816B and NVIDIA Nemotron 3 Ultra 550B on TerminalBench 2.1. It is a general-purpose variant of the firm's chip-design model, with self-reported training cost of about $32 million.
Nanbeige, the foundation model team at Chinese recruitment platform BOSS Zhipin, has released Nanbeige4.3-33B, a dense 33B reasoning model under Apache 2.0 that scores 20 on the Artificial Analysis Intelligence Index v5, matching Qwen3.8 27B (xhigh). It offers a 524K-token context window and runs locally on a laptop, with terminal use and physics reasoning as relative strengths but weaker knowledge benchmarks than Qwen3.8 27B.
Nanbeige, the foundation model team at Chinese recruitment platform BOSS Zhipin, has released Nanbeige4.3-33B under Apache 2.0 — a dense 33B reasoning model scoring 20 on the Artificial Analysis Intelligence Index v5, level with Qwen3.8 27B (xhigh). It offers a 524K-token context window, text-only I/O, and is small enough to run locally on a laptop, with terminal use and physics reasoning as relative strengths but weaker knowledge benchmarks.
IBM has released and open-sourced the Granite 4.2 family of dense, decoder-only reasoning LLMs in 3B, 8B, and 30B sizes. The article details the engineering pipeline used to build these models from Granite-4.1 bases, highlighting techniques like synthetic data generation via Code Alchemy and multi-stage post-training.
IBM releases the open-source Granite 4.1 LLM family (3B/8B/30B), trained on ~15T tokens with long-context extension. The 8B instruct model matches or surpasses the previous 32B MoE architecture in reasoning and tool use, demonstrating that carefully engineered dense models can rival larger sparse architectures.
The Qwen team has open-sourced Qwen3.6-27B, a 27-billion parameter dense multimodal model. It achieves flagship-level performance in agentic coding, surpassing the previous MoE-based Qwen3.5-397B-A17B. Its dense architecture eliminates the need for complex MoE routing, offering developers an ideal solution for top-tier coding capabilities at a deployable scale.