Thinking Machines released Inkling 2 Mini, a 26B total / 5.4B active parameter mixture-of-experts model built on a looped transformer architecture, open-sourced under Apache 2.0 with text and image input and a 1M token context window. It scores 20 on the Artificial Analysis Intelligence Index, matching Qwen3.8 27B and Nanbeige4.3-33B with roughly 5x and 6x fewer active parameters respectively, sits on the Pareto frontier for intelligence versus size, and is small enough to run on a MacBook or DGX Spark.
Active Parameters
激活参数
The parameter count actually engaged per inference in sparse architectures, directly determining inference cost.
Curated Intelligence
Thinking Machines has released Inkling 2 Mini, a 26B-total, 5.4B-active-parameter looped-transformer mixture-of-experts model scoring 20 on the Artificial Analysis Intelligence Index, placing it on the Pareto frontier for intelligence versus active parameters. Released under Apache 2.0 with text and image input and a 1M-token context window, it is small enough to run locally on a MacBook or DGX Spark.
Tencent has released Hunyuan Hy4 Flash, a lightweight model with 224B total parameters and only 12B activated, supporting multimodal input and a 1M context window at roughly 0.26 yuan per task, about 1/20 the cost of Hy4. It topped OpenRouter's daily ranking on launch day with over 50 trillion tokens called, a record for the Hunyuan family, and is now available on Tencent Cloud TokenHub with a limited-time discount.
Thinking Machines, the startup founded by former OpenAI CTO Mira Murati, has released Inkling 2 Small, a 594-billion-parameter mixture-of-experts model that activates only 17 billion parameters per inference. It scores 31 on the Artificial Analysis Intelligence Index, an 18-point gain over its predecessor and roughly on par with Gemini 3.8 Flash, with notable improvements on agentic tasks, though cost per task is six times higher. The model debuts as a proprietary offering on the Tinker platform, with Apache-licensed weights to follow.
Ant Group's Ling team released Ling-3.2-flash, a natively multimodal model built on a 560B-total / ~25B-active MoE architecture with a 1M context window and image, text and video input. In vendor-run tests it beats Claude Opus 5 on OmniDocBench1.5, ZEROBench and Vision2Web, and enters the DesignArena image-to-frontend top ten at 1238 Elo. API pricing is $0.30 per million input tokens and $0.90 per million output tokens, compatible with OpenAI and Anthropic protocols, and free for the first two weeks.
Tencent announced at its Digital Ecosystem Summit that Hunyuan Hy4 is coming soon, claiming broad gains over the preview build in long-horizon and real-world productivity tasks, improved chain-of-thought efficiency, and visual understanding for the first time. A cost-focused Hy4 flash variant will follow shortly. No concrete parameters or benchmark results were disclosed; the August preview had 770B total / 49B active parameters and 1M context.
Mistral AI has released the weights for Mistral Large 4 (ML4), a 1-trillion parameter native multimodal MoE model. It matches leading open-source models on coding agent benchmarks while reducing average token consumption per task by 47%, significantly optimizing inference costs.
Mistral has released the weights for its largest model yet, Large 4 (codenamed 'Le Chonk'), featuring ~1 trillion total parameters and 52 billion active ones under a modified MIT license. While it demonstrates strong coding capabilities in independent benchmarks, it trails behind leading Chinese open-source models and frontier systems from Anthropic and Google on the Artificial Analysis Intelligence Index. Key strengths include high inference speed, lower API costs, and a robust refusal rate for harmful cyber requests.
NVIDIA released the open multimodal model Nemotron 3.5 Super VL, featuring a Mixture-of-Experts (MoE) architecture with 120B total and 12B active parameters. Designed for persistent, multi-step AI agents, it achieves up to 5x higher throughput than comparable open models on a single GPU. The model integrates reasoning, instruction following, tool calling, and multimodal understanding (text, image, document, video), allowing developers to customize behavior via the NeMo toolkit.
NVIDIA released Nemotron 3.5 Super VL, scoring 26.5 on the Artificial Analysis Intelligence Index to become the leading US open-weight model, surpassing competitors from Thinking Machines and Meta. The model doubles its predecessor's intelligence while reducing output tokens per task by ~60%, maintaining near-lossless accuracy in NVFP4 quantization (within 0.1 points of BF16). Concurrently, the new Bolt 74B-A9B achieves performance close to the flagship Ultra model with significantly fewer parameters.