Model Releases & Benchmarks
Granite 4.2 reasoning. IBM released Granite 4.2 in 3B, 8B, and 30B sizes under Apache 2.0. The models have chain-of-thought thinking modes, 512K context, native tool calling, and were post-trained with agentic reinforcement learning.
Thomson Reuters legal model. Thomson Reuters released Thomson-1.0-Small, a law and tax focused model.
Qwen3.8 NVFP4 quant. A fully quantized NVFP4 Qwen3.8-27B checkpoint uses quantization-aware distillation with QUASAR to retain near-BF16 performance, even with all linear layers quantized.
Muse Glimmer benchmark. Muse Glimmer outperformed Qwen3.8 medium in an implicit-knowledge benchmark and took 3-4 hours, compared to 30 hours for Qwen xhigh, suggesting RAG could bring it close to frontier models.
Qwen3.8-Flash-Next local. The upcoming Qwen3.8-Flash-Next is a multimodal MoE with ~125B total params and ~6B active, plus a 51B n-gram table. A 4-bit quant could fit in about 82 GB, making it surprisingly local-friendly.
Hardware & Inference
AMD MI350X kernel. NetraRuntime open-sourced a kernel that runs Qwen3.6-35B-A3B at 78,498 output tokens per second on 8 MI350X GPUs, about 2.16x vLLM throughput.
OpenAI Jalapeño chip. OpenAI's first custom inference chip, Jalapeño, reportedly beats Nvidia Blackwell and Rubin on throughput per watt and latency in SemiAnalysis benchmarks, hitting 1,400 tokens per second per user on GPT-OSS 120B.
DFlash2 speedup. A DFlash2 speculative decoding setup on an RTX 4080 16GB accelerated Qwen3.8-27B from 50.6 to 86.7 tokens per second with identical output quality.
Mac Studio choices. A Mac Studio M5 Ultra 96GB offers double the memory bandwidth of the M5 Max 128GB for about €740 more, but a cost analysis suggests cheaper 24-32GB cards may be more economical for most local inference.
16GB coding test. A heavily quantized Qwen3.8-27B running on a 16GB Quadro correctly implemented a multilayer transfer-matrix method from scratch, though it spent much of the 100-minute run debugging its own validators.
Agents & Frameworks
LangChain agent tooling. LangChain added RubricMiddleware for self-evaluating agents, skills support in deepagents-CLI, and documented loop engineering patterns for stacking agent loops.
LangSmith Engine 2x. LangSmith Engine now identifies impactful agent issues over twice as well on internal benchmarks and is 25% better at writing fixes.
Claude Cowork memory. Anthropic unified memory between Claude chat and Claude Cowork, so Cowork now remembers context from chat and users can view, edit, or delete retained information.
Plan-and-execute agents. LangChain introduced a Plan-and-Execute agent executor that first plans multi-step tasks and then executes them, trading more LLM calls for better long-horizon planning.
Industry & Business
India voice AI. Ringg raised $10M from Peak XV as it builds voice AI agents for Indian enterprises. It already processes 20 million call attempts a month with customers like Flipkart and Cred.
Generalist robotics $3B. Robotics startup Generalist is now valued at $3 billion after an extension bringing its Series B to $600 million. Its Gen 1.5 model lets robots learn tasks from video demonstrations as short as 3 to 12 seconds.
Stability AI funding. Stability AI raised $76 million in Series B from investors including Universal Music, Sony Music, Warner Music, and EA, bringing total funding to $232 million for creative AI tools.
Google legal AI. Google launched Gemini Enterprise for Legal, using MCP connectors to integrate with iManage, NetDocuments, Harvey, and others for contract review, legal research, and regulatory tracking.
Lyft and Klarna agents. Lyft built a self-serve multi-agent customer support platform with LangGraph and LangSmith, while Klarna's LangGraph assistant handles work equivalent to 700 full-time staff.
Safety & Policy
Russia ChatGPT campaign. OpenAI disrupted a Russian covert influence operation that used ChatGPT to generate pro-Kremlin content and fake articles, spread across X, LinkedIn, Facebook, Substack, and Telegram.
EU AI Act compliance. LangChain published guidance on using LangSmith and LangChain OSS to meet EU AI Act high-risk requirements such as logging, monitoring, and human oversight.
That's everything for this day – about a 5-minute read.