Model launchesZhipu’s GLM-5.3 narrows gap with Western models after benchmark showing strengthsChinese startup Zhipu (Z.ai) says its latest GLM-5.3 model outperformed Anthropic’s Mythos 5 on a key cybersecurity benchmark and improved in coding tasks, according to the company.2 min read
Model launchesQwen 3.8 27B: a powerful local multimodal model that over-thinks at default settingsAlibaba's Qwen research lab released Qwen 3.8 27B, an Apache‑2 licensed 27B-parameter multimodal LLM that runs locally as a ~17GB file and shows improved benchmarks over prior Qwen versions.5 min read
Model launchesDeepSeek V4 Flash shows strong capabilities but limited reliability in multi-tool testsDeepSeek’s V4 Flash has drawn rapid developer adoption since its public beta, but independent multi-agent tests found it completed just 53.8% of complex workflows.4 min read
Model launchesZ.ai’s GLM-5.3 narrows the gap with smaller model size; faster release cycles and dual‑use cyber risksZ.ai published GLM-5.3, a ~750B-parameter model released first to its coding plan with API and open weights coming soon; the company credits extended post‑training (RL‑dominated) for large benchmark gains that in many tests match or exceed some leading Western public models.5 min read
Model launchesGoogle appears to prioritize volume and price over frontier model ambitionsGoogle released Gemini 3.7 Flash just three weeks after 3.6 Flash, cutting the new model's price in half and emphasizing code- and agent-focused use cases.2 min read
Model launchesMeta launches open-weight model Glimmer as Zuckerberg champions 'AI for everyone' with caveatsMeta released Glimmer, an open-weight AI model available for download and local use, while keeping its more capable Muse Spark model behind proprietary APIs.2 min read
Model launchesApple develops China-specific large language model for iPhone AI featuresApple has trained its own Chinese-language large language model for use with iPhone AI features, developed with support from Alibaba Group, according to sources.2 min read
Model launchesOpenAI unveils ChatGPT 'Ultrafast' mode to accelerate GPT-5.6 SolOpenAI has introduced an "Ultrafast" mode for ChatGPT that accelerates the GPT-5.6 Sol model, claiming a 14-fold increase in processing speed and output rates up to 750 tokens per minute.2 min read
Model launchesUltrafast model with Cerebras accelerator: up to 750 tokens/second performanceThe Ultrafast language model runs on Cerebras hardware and can produce up to 750 tokens per second, providing low-latency intelligence for real-time speech processing, customer service, commerce, coding, design, financial research, and security response.1 min read
Model launchesDeepSeek launches V4‑Pro GA and open-source Harness while shifting API to peak/off‑peak pricingDeepSeek released the official DeepSeek‑V4‑Pro model and an MIT‑licensed open‑source agent harness called DeepSeek Harness (dsh) on Aug.5 min read
Model launchesWriter launches Palmyra X6 and harness upgrades to cut token costsWriter introduced Palmyra X6, a post-training variant of Z.ai’s open-source GLM-5.2, together with upgrades to its agentic harness intended to reduce token costs for customers.3 min read
Model launchesGPT‑5.6 cuts agent costs with model selection and new API primitivesThe GPT‑5.6 model family improves agent performance while substantially reducing inference costs by enabling smaller models to handle longer-horizon tasks and by introducing new Responses API primitives.4 min read
