A chart from whatmodel.sh, an open-source AI leaderboard, shows Chinese open-weight text models trail US closed frontier models by about seven months, based on 22,113 results across 987 benchmarks from 1,011 sources.
LFM2.5-24B-A2B is an open-weight Mixture-of-Experts model from Liquid AI with 24B total and about 2B active parameters, designed to run fast, private agent tasks locally on a laptop. Weights are available on Hugging Face.
Z.ai's GLM-Legal, built on GLM-5.3-Flash, is a legal AI model scoring 19.2% all-pass on Harvey's Legal Agent Benchmark, nearly 3x its base, at $0.15 per test, by application.
OrcaRouter compares two text-to-speech systems, Fish Audio Drama 3 Preview and Qwen Audio 3.0 TTS, on voice quality and expressiveness.
QwenGyre is an elastic online reinforcement learning framework from Alibaba's Qwen team for training extreme-long-horizon coding agents through unmodified black-box harnesses, targeting rollouts that run for hours.
MiniCPM-V 4.7 is the latest in OpenBMB's vision-language series, including a 35B-A3B variant that is its first MoE model, plus a 1B version and GGUF quantizations. It handles single-image, multi-image and video understanding.
Boole-1.7B is a 1-bit version of Qwen3-1.7B with 247 MB packed weights, 13.9x smaller than BF16 and scoring 85.9% of the BF16 model across ten tasks. It can be run locally.
An independent developer continued work on converting dense LLMs into Mixture-of-Experts models, releasing Qwen2.5-1.5B-MoE-A0.9B along with GGUF versions for local use.
A project repurposes Google's EmbeddingGemma 2 embedding model into a speech-to-text system, with the write-up and demo shared on Hacker News.
Shows a small active-parameter MoE can run capable agent workloads privately on-device.
LFM2.5-24B-A2B is an open-weight Mixture-of-Experts model from Liquid AI with 24B total and about 2B active parameters, designed to run fast, private agent tasks locally on a laptop. Weights are available on Hugging Face.
It quantifies the open-versus-closed gap that shapes policy and competition in AI.
A chart from whatmodel.sh, an open-source AI leaderboard, shows Chinese open-weight text models trail US closed frontier models by about seven months, based on 22,113 results across 987 benchmarks from 1,011 sources.
FLock.io released an edge decision model built on Qwen aimed at enhanced data privacy, claiming it outperforms larger LLMs on selected tasks.
A 9B-active 74B MoE squeezed to 24GB would make near-frontier quality run on a single GPU.
A Reddit post says Nvidia's Nemotron Bolt 3.5 is a 74B-parameter, 9B-active MoE model quantized to 24GB using 2-bit quantization-aware training for speed.
Frames post-training — not compute — as the lever behind China's model progress.
BigGo Finance argues Z.ai's GLM 5.3 shows post-training techniques, rather than raw compute, have become China's key competitive advantage in AI.
A specialized legal agent model nearly tripling its base score shows domain tuning still beats general scaling.
Z.ai's GLM-Legal, built on GLM-5.3-Flash, is a legal AI model scoring 19.2% all-pass on Harvey's Legal Agent Benchmark, nearly 3x its base, at $0.15 per test, by application.
OrcaRouter contrasts a quietly published repo, HiRO-ACEv2, against the full launch of Qwen 3.8, weighing the two model efforts. The source is largely commentary without detailed benchmark numbers.
A competitive open-weight Indic voice model broadens multilingual speech AI options.
An X post announces Maya-2, billed as the world's best Indic voice AI model with fully open weights, claiming it outperforms ElevenLabs v3, Cartesia Sonic 3.5 and Bulbul v3 across Indian languages; free to builders soon.
An enthusiast ran a 27 billion parameter Qwen model on a jailbroken PlayStation 5's GPU at about 20 tokens per second, demonstrating local LLM inference on console hardware.
Frames the open-weights safety debate now driving policy.
MIT Technology Review Explains piece examines the benefits of open-weight models against risks, citing an Anthropic report arguing GLM-5.3 poses a major cyberattack risk.