StepFun's Step 3.7 Flash is now available in the Baseten Model Library in a hardware-efficient configuration. This 198-billion-parameter sparse MoE vision-language model supports multimodal input, a 256k context window, and flexible reasoning for agentic workflows. FP8 quantization enables deployment on 4×H100 GPUs, significantly reducing cost.

ElevenLabs officially launches in Canada, appoints Max Lemmens as General Manager, plans to double the team and open a Toronto office. The company already has 30,000 users in Canada, including major telcos and media outlets. Canada's bilingualism and key industries present a strong opportunity for voice AI.

Gumloop achieved 7x growth in open-weight model agent chats within three weeks by partnering with Fireworks AI, saving up to 72% on costs while maintaining quality.

Cohere proposes Dynamic Speculative Decoding (DSD) that adaptively adjusts the number of draft tokens based on hardware constraints to accelerate LLM inference. It addresses the performance degradation of fixed-k speculative decoding at high batch sizes, validated on dense and MoE models, and is compatible with vLLM's async scheduling and CUDA Graph optimizations.

Cerebras collaborates with Upstage to bring ultra-fast inference (up to 2,000 tokens/sec) on Solar 31B model, targeting enterprise AI applications in Korea and beyond.

Semantic search returns results that are close in meaning but may miss unstated context, leading agents to act on incorrect information. Pinecone's new text match filters, now in public preview, allow lexical filtering to scope search results without requiring pre-labeled metadata for every potential query dimension. This reduces errors in agentic pipelines where a bad retrieval can trigger a chain of wasted tool calls.

Cerebras has redesigned its technical interview process to allow candidates to use AI, focusing on problem framing, verification, judgment, communication, and ownership. AI collaboration is evaluated as a skill, and verification becomes a key signal.

People have many hard questions about AI. Anthropic, as a Public Benefit Corporation, aims to address these concerns transparently. They have surveyed thousands and now invite the public to submit their toughest questions.

LangChain has tuned its Deep Agents harness for NVIDIA Nemotron 3 Ultra, achieving benchmark-leading agent performance among open models at 10x lower cost than closed alternatives. The tuned harness is available in LangChain Deep Agents, and Nemotron 3 Ultra runs on Fireworks with day-zero support.

Former Federal Reserve Chair Ben Bernanke joins Anthropic's Long-Term Benefit Trust, bringing expertise in economics to guide the company's AI governance and impact analysis.

Mistral Studio introduces version control for prompts and skills, turning them into governed assets with immutable versions, rollback, clear ownership, and audit logs. It enables fast iteration and controlled deployment, solving the problem of scattered prompts across enterprises.

Mistral Studio introduces version control for prompts and skills, turning them into governed assets with immutable versions, rollback, clear ownership, and audit logs. It enables fast iteration and controlled deployment, solving the problem of scattered prompts across enterprises.

Anthropic introduces a beta reflection feature to help users track and visualize their Claude usage patterns, align time spent with personal goals, and build AI skills through the 4D AI Fluency Framework. Available for Free, Pro, and Max users with memory turned on.

Using GLM 5.2 Fast via FireConnect on Claude Code, the author designed, planned, and implemented a GPU scheduler reclaim feature in four days at a cost of $218 in inference tokens—a task normally scoped at one engineer-month. The article highlights how fast inference speed (400 tokens/sec) eliminated context switching, low token cost removed usage anxiety, and model quality handled complex concurrent logic.

This article showcases three multimodal applications built with Gemma 4 on Cerebras, including document processing, image understanding, and video analysis. The Gemma 4 31B model achieves ~2,300 tok/s on Cerebras, enabling real-time multimodal interactions.

LlamaIndex ships multiple product updates in a busy June, including Retrieval Harness for agent file-system tools, LiteParse markdown support, MCP endpoint restructuring, Cost Optimizer improvements, enterprise Usage Tags and User Metadata. Community highlights include a hands-on build with LiteParse and LanceDB, an n8n community node, and several event talks. AI news covers Anthropic's Fable 5 extension, OpenAI DevDay 2026 applications, and more.

Cursor and SpaceXAI release Grok 4.5, the most intelligent model yet, designed for more than software engineering. It handles complex, long-running tasks across data science, finance, legal work, and more. Trained as a mixture-of-experts model on trillions of tokens of Cursor interaction data, it uses reinforcement learning on difficult problems. Available now on Cursor with significant usage included, doubled for the first week.

Build frontier AI agents with NVIDIA Nemotron 3 Ultra and LangChain Deep Agents on Baseten—top open-model agent accuracy at ~10× lower cost.

Robostral Navigate is an 8B model enabling robots to autonomously navigate complex environments using only a single RGB camera, achieving 76.6% success on unseen R2R-CE benchmarks—outperforming multi-sensor approaches. Built with simulation-trained data and token-efficient techniques, it generalizes across robot types and adapts to real-world obstacles. The model combines pointing-based navigation with reinforcement learning for continuous improvement, paving the way for unified embodied AI.

Robostral Navigate is an 8B model enabling robots to autonomously navigate complex environments using only a single RGB camera, achieving 76.6% success on unseen R2R-CE benchmarks—outperforming multi-sensor approaches. Built with simulation-trained data and token-efficient techniques, it generalizes across robot types and adapts to real-world obstacles. The model combines pointing-based navigation with reinforcement learning for continuous improvement, paving the way for unified embodied AI.

GLM 5.2 Fast is available on Fireworks, offering Opus-level intelligence at open-source rates. It runs 2-3x faster than Standard on shared serverless infrastructure, with a full 1M-token context window, aggressive prompt caching, and structured output support. Fireworks achieved 446 tok/sec and optimized the model's MoE and sparse attention architectures for real-world workloads.

This article introduces how Unstructured Platform handles over 60 unstructured data formats using a multi-layered parsing strategy, including rule-based parsers, OCR, and VLMs. It also discusses the trade-off in MCP server tool count and how to optimize context window usage.

Since 2025, Alberta's Ministry of Technology and Innovation used Claude Code with Opus and Sonnet to scan 466M lines of code in 20 hours, found and fixed vulnerabilities, and built tools. It would have taken 6.5 years traditionally.

Global AI spend reaches $1.5 trillion in 2025, yet only 39% of organizations trace it to EBIT impact. Cursor launches the CFO Council to develop shared frameworks for measuring and managing AI economics. Data shows highly uneven returns and costs, with top users getting 46x more output. The council will meet quarterly to establish benchmarks and best practices.

Leanstral 1.5, a free Apache-2.0 licensed model with 6B active parameters, delivers a major performance upgrade in formal verification, saturating miniF2F, solving 587/672 PutnamBench problems, and achieving state-of-the-art results on FATE-H (87%) and FATE-X (34%). Trained through mid-training, supervised fine-tuning, and reinforcement learning with CISPO, it excels in agentic proof engineering and real-world code verification, uncovering 5 previously unknown bugs across 57 repositories tested. Fully open-sourced and available via Hugging Face and a free API, Leanstral 1.5 is now accessible for practical proof engineering in Lean 4.

Anthropic has re-deployed Claude Fable 5 globally and provided additional details on its cybersecurity classifiers and a proposed AI jailbreak severity framework. The classifiers categorize cybersecurity activities into four tiers: prohibited, high-risk dual-use, low-risk dual-use, and benign, with corresponding blocking or monitoring policies. The jailbreak severity framework aims to establish a common language for discussing risks across industry, academia, and government.

Object Detection (OD) remains a critical foundation for document transformation pipelines, even with the rise of VLMs. This article explains why fine-tuning OD models is far from plug-and-play, covering challenges such as training framework bugs, inconsistent label philosophies across datasets, and catastrophic forgetting. Unstructured's approach, building on IBM Heron, achieved significant accuracy gains through careful fine-tuning.

H100, H200, and B200 GPUs each make different tradeoffs in memory, compute, and cost. Learn which one fits your model size, traffic, and budget.

Flows Agent is a conversational AI assistant in ElevenCreative Flows that builds and runs creative pipelines based on user descriptions, enabling rapid iteration, cost control, background execution, and multiplayer collaboration.

ElevenLabs introduces Procedures in ElevenAgents, a feature that enables defining structured or free-form instructions for AI agents, improving efficiency and consistency in tasks like refunds, billing, and troubleshooting. Supports importing existing SOPs.