AI Briefing#llm
The latest AI and developer news about #llm, with the original source and a short summary.
The latest AI and developer news about #llm, with the original source and a short summary.
AI automation boosted travel product registration speed from 20 per day to over 5,000.
Ternary Bonsai has released 8B, 4B, and 1.7B models using 1.58-bit ternary weights.
Parcae uses a looped recurrent structure to make 770M aim for 1.3B Transformer-level performance.
Alibaba has open-sourced the Qwen3.6-35B-A3B model, which features agentic programming capabilities.
This is a real-usage report that 4.7 violates CLAUDE.md prohibition rules more often than 4.6.
Qwen3.6-35B-A3B has been released, achieving large-model-class coding and multimodal performance with just 3 billion active parameters.
Mozilla's Thunderbird team has unveiled Thunderbolt, an open-source AI client that emphasizes data sovereignty and model independence.
A new Android CLI has been released that integrates with AI agents to speed up Android app development by up to 3x.
PrismML has released Ternary Bonsai, a family of 8B, 4B, and 1.7B LLMs using 1.58-bit ternary weights.
Anthropic has launched Claude Opus 4.7, a flagship model with self-verification capabilities and high-resolution vision.
Google's Gemma 4 model supports on-device AI inference that runs on iPhone without an internet connection.
Alibaba's Qwen team has released Qwen3.6-35B-A3B, an open-source MoE-based model with significantly improved agentic coding performance.
The Claude Opus 4.7 model card specifies synthetic data generated by other models.
Role, process, standards, and honesty—not 'unlocking'—determine output quality.
llm-anthropic 0.25 supports Claude Opus 4.7 and new thinking options.
Even the same input can result in actual API costs differing by several times depending on the tokenizer.
Assume LLMs hallucinate, loop, and ignore tools — wrap them with subagents and scripts.
Real use cases exist, but adoption remains limited due to security, cost, and stability issues.
Android CLI, skills, and Knowledge Base accelerate Android agent development.
Swapped Qwen 1.5B's decoder head with a CHIP-8 opcode head and ran it directly.
A benchmark running MiniMax M2.7 AWQ on 8x MI50 32GB hit a peak of 64 tok/s.
Claude Opus 4.7 removes manual thinking settings and supports only adaptive thinking.
Manual extended thinking has disappeared in Opus 4.7, leaving only adaptive thinking.
It compresses the context of each LLM call into a Manifest to auto-bootstrap the evaluation stack.