LLMs

Hugging Face and NVIDIA report that a single Nemotron model family achieved gold-level results in both the International Olympiad in Informatics and the International Mathematical Olympiad. The achievement highlights the effectiveness of targeted fine-tuning strategies for high-stakes competitive programming and advanced mathematics. This demonstrates the model's capacity to handle complex, logic-intensive tasks when properly optimized.

• Nemotron model family secured gold-level performance in IOI and IMO 2026.
• Single model architecture handles both coding and advanced math competitions.
• Results underscore the value of specialized fine-tuning for elite benchmarks.
• Validates end-to-end optimization for logic-heavy, multi-modal reasoning tasks.

Amazon Bedrock now accepts the reasoning.summary parameter for OpenAI models through the Responses API. This feature returns a human-readable summary of the model's internal reasoning process alongside the final answer. Developers can use this output to debug applications, evaluate complex tasks like coding, and provide transparency to end users.

• Enable reasoning.summary in the Responses API to get a summary array in the output.
• Use the reasoning summary to debug logic errors in coding or multi-step analysis tasks.
• Improves model transparency by showing users how the AI approached a complex request.
• Available for all OpenAI models on Bedrock in supported regions and inference modes.

Safety-aligned models frequently bypass refusal mechanisms when harmful requests are embedded in role-play or narrative contexts. A new benchmark called GUISE measures this vulnerability across English, Modern Chinese, and Classical Chinese registers. Attack success rates on Qwen3-1.7B are strikingly high, reaching 95.7% in Classical Chinese. The study also introduces a stricter evaluation criterion that counts warn-then-answer responses as successful attacks.

• Narrative wrapping is a potent jailbreak vector for LLMs across multiple languages
• Qwen3-1.7B shows 89-95% attack success depending on the language register
• Classical Chinese wrappers are particularly effective against safety filters
• GUISE benchmark standardizes testing with parallel requests and held-out wrapper types
• Warn-then-answer outputs now count as attack successes in this evaluation

Amazon Bedrock has made Z.ai’s GLM 5.3 model generally available, targeting agentic coding and long-horizon software engineering tasks. This mixture-of-experts model features 753 billion total parameters with approximately 40 billion active per token, leveraging scaled post-training improvements. It supports a 1-million-token context window and up to 128K output tokens, with reasoning capabilities that allow users to adjust effort levels for latency and cost trade-offs.

• GLM 5.3 is now GA on Bedrock, offering AWS security and compliance for open-weight models.
• MoE architecture delivers 753B total params with ~40B active, optimizing inference efficiency.
• Supports 1M token context and 128K output, suitable for complex, long-horizon engineering workflows.
• Reasoning is always enabled with selectable effort levels to balance performance against latency.
• Bedrock integration includes explicit prompt caching with configurable cache points for system prompts.