MarkTechPost
·
1 day ago
Feyn AI released SQRL, a family of text-to-SQL models that inspect databases before generating queries to resolve ambiguity and avoid logically incorrect results. SQRL-35B-A3B achieves 70.6% execution accuracy on the BIRD Dev benchmark, outperforming Claude Opus 4.6 at 68.77%, with smaller 4B and 9B variants also available on Hugging Face. The models are trained to decide when inspection is needed, allowing easy questions to be answered immediately while complex queries benefit from read-only database probes before final answer generation.
MarkTechPost
·
1 day ago
● 5 sources
Alibaba previewed Qwen3.8-Max-Preview, described as a 2.4 trillion-parameter multimodal model, during Shanghai's World AI Conference on July 19, 2026, two days after Moonshot AI released its 2.8 trillion-parameter Kimi K3 model. The preview is available now at 10% standard pricing through Alibaba's Token Plan subscription, but the full model card, benchmark table, and open-weight release date remain unpublished. Without disclosure of active parameters per token—the actual number of parameters used per inference—the real serving cost and practical value of the model cannot be assessed, leaving developers dependent on unverified performance claims.
The Neuron
·
1 day ago
● 10 sources
Kimi K3 is an open-source model that has pushed the capabilities of accessible language models closer to frontier performance levels. The model demonstrates competitive performance across benchmarks typically dominated by proprietary systems. This development increases competition in the open-model space and provides researchers and developers with a more capable freely available alternative to commercial offerings.
Sakana AI
● 2 sources
Researchers introduced Text-to-LoRA, a hypernetwork that generates task-specific LoRA adapters for large language models by taking text descriptions as input rather than requiring expensive fine-tuning. The system can encode hundreds of existing LoRA adapters and generalize to unseen tasks, with a reference implementation released on GitHub using 7B models. This approach reduces computational barriers and technical expertise needed to customize foundation models for specific applications using plain language descriptions.
Sakana AI
● 2 sources
Sakana AI published a paper in Nature Machine Intelligence describing Evolutionary Model Merge, a method that uses evolutionary algorithms to automatically combine existing open-source AI models into new foundation models optimized for specific tasks. The company created three Japanese foundation models (EvoLLM-JP, EvoVLM-JP, and EvoSDXL-JP) where their 7B-parameter Japanese math LLM matched the performance of previous 70B-parameter models without requiring gradient-based training. This approach enables automated foundation model development with minimal compute resources by systematically discovering effective model combinations that human experts might not intuitively discover.
Sakana AI
● 2 sources
Sakana AI developed AB-MCTS, an inference-time scaling algorithm that enables multiple frontier AI models to cooperate and perform trial-and-error reasoning. A combination of o4-mini, Gemini-2.5-Pro, and DeepSeek-R1-0528 using Multi-LLM AB-MCTS achieved significantly higher performance on the ARC-AGI-2 benchmark compared to any individual model. This approach leverages the diverse strengths and weaknesses of different models to solve problems that would be insurmountable for any single model.
Sakana AI
● 2 sources
Sakana AI presented a paper at GECCO'25 proposing M2N2, a method that uses evolutionary algorithms to merge AI models by automatically determining how to partition and combine them rather than requiring manual definition. The approach evolved an MNIST classifier from random networks and successfully merged a math specialist LLM with an agentic specialist LLM to handle both tasks better than existing methods. This evolutionary model fusion approach enables more flexible combination of specialized models while avoiding catastrophic forgetting seen in traditional fine-tuning.
Sakana AI
Researchers reported findings in March 2025 that their AI CUDA Engineer evaluation had benchmark vulnerabilities that allowed artificial optimization without genuine performance gains. They developed a more robust benchmark called robust-kbench that eliminated these exploitable loopholes, and re-evaluation showed LLM-based CUDA kernel optimization achieved an average speedup of 1.49 times instead of the originally reported 3.13 times. The corrected benchmark provides a more reliable foundation for evaluating AI-assisted code optimization going forward.
Sakana AI
● 2 sources
Sakana AI released ShinkaEvolve, a framework that uses LLMs to evolve and discover new algorithms with dramatically improved sample efficiency compared to prior evolutionary approaches. The system discovered a state-of-the-art Circle Packing solution using only 150 samples, designed an effective math competition agent scaffold in 75 generations, and found a novel loss function for Mixture-of-Experts models after 30 generations. The open-source framework enables researchers and engineers to use evolutionary AI discovery as a practical tool for optimizing algorithms and training strategies across multiple domains.
Sakana AI
● 2 sources
Researchers introduced Petri Dish Neural Cellular Automata (PD-NCA), an artificial life system where neural cellular automata learn continuously during simulation rather than using fixed parameters. The system enables multiple NCA to self-replicate and compete through ongoing gradient descent optimization. Complex emergent behaviors including cyclic dynamics, territorial defense, and spontaneous cooperation arise from individual organisms constantly adapting to out-compete neighbors.
TheSequence
·
2 days ago
● 10 sources
Multiple AI labs released competing model architectures this week: Thinking Machines open-sourced Inkling (975B parameters, open weights), Moonshot AI launched Kimi K3 (2.8T parameters), and PrismML released Bonsai 27B (compressed to 3.9GB for smartphones). OpenAI introduced GPT-Red, an automated red-teaming system that successfully compromised GPT-5.1 in 84% of test scenarios through self-play training. The week's developments reflect a shift from centralized AI concentration toward distributed, adapted, and locally-deployable models, with geopolitical implications as China's Xi Jinping promoted open-source AI as a global public good at the Shanghai World AI Conference.
Exponential View
·
2 days ago
● 6 sources
Moonshot AI's Kimi K3 model performs comparably to Claude Opus and GPT 5.6 but costs 24 times more than DeepSeek V4 Pro on a per-token basis. The emergence of competitive open-weight models will increase token demand and push revenue toward infrastructure providers rather than model companies. As AI capabilities become more commoditized through open models, ensuring free learning from lawfully accessible material—similar to copyright frameworks that historically enabled knowledge compounding—will be critical for competitive positioning.
MarkTechPost
·
2 days ago
● 6 sources
Three Chinese AI labs released large open-weight Mixture-of-Experts models: Moonshot AI's Kimi K3 (2.8 trillion parameters), DeepSeek V4 Pro (1.6 trillion), and Zhipu AI's GLM-5.2 (744 billion). Kimi K3 scores 57 on the Artificial Analysis Intelligence Index and ranks third overall, while DeepSeek V4 Pro costs $0.04 per task and has weights available immediately under MIT license, whereas K3 remains API-only until July 27, 2026. Teams choosing models must trade off capability, cost (ranging from $0.18 to $3.00 per million input tokens), and availability of downloadable weights for self-hosting.
