DeepSeek, MiniMax Complete China's AI Export Registration
State media say DeepSeek and MiniMax completed China's new AI export registration, and MiniMax confirmed its own in a filing. Experts say it is not an exemption.

State media say DeepSeek and MiniMax completed China's new AI export registration, and MiniMax confirmed its own in a filing. Experts say it is not an exemption.

OpenBMB's Apache 2.0 MiniCPM-OCR 1.0 reads scans, tables and formulas at 1.2 pages per second on one GPU, and a 2B Jev model scores OCR output for review.

On October 29, 2026, Alibaba Cloud's CTO reportedly said "Qwen4 next week" in Amsterdam. What the slide showed, what is confirmed, and the Qwen Code clue.

Fish Audio Drama 3 is open-weight text-to-speech you direct with bracketed stage notes. Its scores, 48 GB GPU need, license terms and $10 launch API price.

Baseten's MIT-licensed joey-1 returns calibrated yes/no, choice and score probabilities in one pass, at $0.01 per million input tokens. Claims and limits.

Pika H3 Flex is a MiniMax H3-based video model with open weights, up to 12 references, five audio stems and hosted 4K upscaling. Specs, prices and limits.

ServiceNow's MIT-licensed Apriel Voice 27B Preview takes speech in and returns text and tool calls. Its benchmarks, hardware needs and limits, explained.

Mistral Large 4, DeepSeek-V4.1-Flash and Qwen3.8-2.4T-A95B compared on cost per task, license, hardware and fit, from model cards and Artificial Analysis.

Open-weight models let you download and run the model; closed models run on the developer's service. How they differ, what licenses allow, and how to choose.

IBM's Granite Switch 4.1.1 adds multilingual speech input and new Decision adapter weights, while the base model stays untouched.

Opinion: Mistral says it will keep updating Mistral Large 4's cyber guardrails. Those fixes can reach its API, but downloaded open weights stay as shipped.

Reactions to Mistral Large 4's open weights: Artificial Analysis says it narrowly leads non-Chinese open models; HN debated the license and $0.70/$2.10 price.

DiffuSpace's Clover-35B-A3B-Preview is an Apache 2.0 diffusion language model built from Qwen3.6-35B-A3B. How it works, its claimed scores and how to run it.

Mistral Large 4 weights are on Hugging Face in FP8 and NVFP4, with GGUF quants. The license limits, the 1T MoE design, reported benchmarks and hardware needs.

Two Swift packages by Mattt, on the Hugging Face blog, run Parakeet and Qwen3-ASR behind Apple's speech API and read decisions from model probabilities.

Particle post-trained DeepSeek V4.1 Flash with RL and LoRA. It says Particle-1 uses about 45% fewer tokens with comparable accuracy, at the same price.

A developer compiled a RISC-V CPU into Qwen3.5-4B's weights. Reactions from the Qwen team, X users who tested it, and skeptics on Hacker News.

Irregular says Zhipu's open-weight GLM-5.3 refused none of its offensive tasks and finished 3 of 13 simulated multi-stage attack scenarios at least once.

SenseTime's SenseNova 6.8 Code2Video is an open 27B model that makes short videos by writing and rendering code. What it ships, how it runs, and its limits.

A viral X recipe pushes one neuron in Alibaba's open Qwen3.8 27B to make it act like a villain. What single-neuron steering is and what it does not prove.
