THIS WEEK SHOULD BE EXCITING.me: every week, forever.

↓.

DeepSeek-V3.1 ⮕ looks like they’ve been tinkering

COLLECTION: TR:

ASearcher gives open LLMs full “Search Intelligence,” fixing ≤10-turn RL and weak fact-checking via 128-turn async RL (40+ tool calls, 150k+ tokens), synthetic QA gen (14k→134k verified hard QAs, 25.6k tool-required), and a minimal-tool unified agent. Results: 14B: RAG F1 60.0

COLLECTION.

GLiNER-decoder is an encoder–decoder model for NER that uses span pooling and autoregressive type generation. - Removes GLiNER’s ~30-type cap, label-order dependence, and fixed ontology.- A single encoder pass produces span vectors → decoder prompts generate types (supports

My technical deep dive on GLiClass-V3 ⮕ (papers out) . Training pipeline: .- Pre-train on 1.2M NLI/classification mix.- PPO-adapted multi-label RL mid-train (few-shot & feedback learning).- LoRA post-train with logic/NLI + pattern-focused streams (high-rank LoRA → edge

COLLECTION: PAPER:

InclusionAI releases GroveMoE . GroveMoE uses heterogeneous + adjugate experts (ARM’s big.LITTLE-inspired). Experts are grouped, each sharing an adjugate expert computed once, so compute scales with token complexity (3.14–3.28B active per token). Models.- GroveMoE-Base (33B,

MODELS: SETS: RELEASE:

NVIDIA Nemotron-Nano v2. Models: 12B Base, 9B Reasoning, 9B Base.- Arch: Hybrid Mamba2–Transformer (128K ctx, 4 attn layers).- Training: 10.6T tokens (3.5T synthetic from DeepSeek, Qwen, Nemotron-4, phi-4, etc.).- 15 natural languages + 43 programming languages.- Datasets:

𝐓𝐇𝐈𝐒 𝐈𝐒 𝐒𝐎𝐌𝐄 𝐆𝐑𝐄𝐀𝐓 𝐖𝐎𝐑𝐊 . …𝐃𝐎𝐍’𝐓 𝐒𝐋𝐄𝐄𝐏

Cool list… but I’ve been tracking this space closely. Might be time I share my own ranking of China’s open model labs.

SET: CANARY: PARAKEET:

NVIDIA DROP: multilingual ASR + AST stack. - Granary: ~1M hours of ASR/AST data across 25 European languages, built with the NeMo SDP pipeline from YODAS, MOSEL, YTC, and LibriLight. Largest open-source EU speech dataset, backbone for Canary and Parakeet training. - Canary-1b-v2:

RT @gm8xx8: Kimina-Prover-RL is a DeepSeek-R1-style RL pipeline for Lean 4 theorem proving from Project Numina and Kimi, built on Verl. Two….

☺︎