Google DeepMind released the Gemma 4 family of open-weights models ranging from 2.3B to 31B parameters, featuring up to 256K token context windows and native support for text, image, video, and audio inputs. The flagship 31B model scores 85.2% on MMLU Pro and 89.2% on AIME 2026, with a smaller 26B MoE variant requiring only 3.8B active parameters for faster inference.
Microsoft AI released three foundational models on April 2: MAI-Transcribe-1 for speech-to-text across 25 languages, MAI-Voice-1 for audio generation, and MAI-Image-2 for video generation. The company positions these models as cheaper alternatives to Google and OpenAI offerings. Models are available on Microsoft Foundry with pricing starting at $0.36 per hour for transcription.
NVIDIA has optimized Google's Gemma 4 models for local deployment on RTX and Spark platforms, targeting the emerging wave of on-device agentic AI. The optimization enables small, efficient models to access real-time local context for autonomous decision-making without cloud dependency.
Microsoft has released MAI-Transcribe-1, a speech-to-text model that achieves the lowest word error rate on the FLEURS benchmark across 25 languages, outperforming Whisper-large-V3, GPT-Transcribe, and Gemini 3.1 Flash-Lite. The model runs 2.5 times faster than Microsoft's previous Azure Fast offering and costs $0.36 per audio hour.
Google DeepMind released Gemma 4, a family of four open models ranging from 2B to 31B parameters, all licensed under Apache 2.0. The 31B dense model ranks #3 on Arena AI's text leaderboard and the 26B mixture-of-experts variant ranks #6, outperforming closed models significantly larger in size.
MLCommons published MLPerf Inference v6.0 results on April 1, 2026, with Nvidia, AMD, and Intel each claiming top spots in different configurations. Nvidia's 288-GPU GB300-NVL72 system achieved 2.49 million tokens per second on DeepSeek-R1, while AMD's MI355X crossed one million tokens per second for the first time. Direct comparisons remain difficult as each chipmaker targets different market segments and benchmarks.
Alibaba has released Qwen3.6-Plus, its third proprietary AI model in days, featuring a 1 million token context window available via Alibaba Cloud Model Studio API. The model claims improved agentic coding capabilities and partially outperforms Anthropic's Claude 4.5 Opus in Alibaba-conducted benchmarks, though trails Claude 4.6 Opus released in December 2025.
Amazon Bedrock AgentCore Runtime now offers managed session storage to persist agent filesystem state across stop/resume cycles and InvokeAgentRuntimeCommand for executing shell commands directly within agent microVMs. The features address two core challenges in production agent workflows: ephemeral filesystems that reset between sessions and the inability to execute deterministic operations without routing them through LLMs.
Alibaba's Qwen division released Qwen 3.6 Plus on April 2, 2026, offering free access to a model with a 1,000,000 token context window. The model combines linear attention with sparse mixture-of-experts routing and achieves a 78.8 score on SWE-bench Verified for software engineering tasks.
OpenAI President Greg Brockman stated that GPT reasoning models have 'line of sight' to AGI and represents a settled debate on whether text-based models can achieve general intelligence. The company is prioritizing this approach over multimodal world models like Sora, which Brockman views as 'a different branch of the tech tree.' The stance contradicts prominent AI researchers including Yann LeCun and Demis Hassabis, who argue LLMs alone are insufficient for human-level intelligence.
Google has unveiled TurboQuant, a KV cache quantization technology that claims to reduce memory consumption during AI inference by up to 6x by compressing data from 16-bit precision to as low as 2.5 bits. While the compression technique delivers meaningful efficiency gains for inference providers, it is unlikely to resolve the DRAM shortage that has driven memory prices to record highs, as expanding context windows offset memory savings.
Claude Code will automatically approve denied commands—like curl—if preceded by 50 or more chained subcommands, according to security firm Adversa. The vulnerability stems from a hard-coded MAX_SUBCOMMANDS_FOR_SECURITY_CHECK limit set to 50 in the source code, after which the system falls back to requesting user permission rather than enforcing deny rules.
Amazon Web Services has detailed how its Nova Act browser automation SDK can streamline competitive price intelligence workflows. The service enables developers to build agents that navigate websites, extract pricing data using natural language instructions, and run parallel monitoring across multiple competitor sites—addressing manual processes that consume hours daily and delay pricing decisions.
Z.ai has released GLM-5V Turbo, a native multimodal foundation model designed for vision-based coding and agent-driven tasks. The model supports image, video, and text inputs with a 202,752 token context window, priced at $1.20 per million input tokens and $4 per million output tokens.
A Google Deepmind paper introduces the first systematic framework for 'AI agent traps'—attacks that exploit autonomous agents' vulnerabilities to external tools and internet access. The researchers identify six attack categories targeting perception, reasoning, memory, actions, multi-agent networks, and human supervisors, with proof-of-concept demonstrations for each.
H Company unveiled Holo3, a computer use model that scores 78.85% on the OSWorld-Verified benchmark—the highest on the leading desktop automation benchmark. The model achieves this with only 10B active parameters (122B total), positioning it as a lower-cost alternative to proprietary models like GPT 5.4 and Opus 4.6.
Arcee AI has released Trinity Large Thinking, an open-source reasoning model featuring a 262,144 token context window. The model is priced at $0.25 per million input tokens and $0.90 per million output tokens, with free access available through OpenRouter for the first five days.
Anthropic's Claude Code version 2.1.88 release accidentally included a source map exposing over 512,000 lines of code and 2,000 TypeScript files. Analysis of the leaked codebase by security researchers reveals evidence of a planned 'Proactive' mode that would execute coding tasks without explicit user prompts, plus potential crypto-based autonomous payment systems.
Elgato has released Stream Deck 7.4 with Model Context Protocol (MCP) support, enabling AI assistants like Claude and ChatGPT to find and activate Stream Deck actions on your behalf. Users can now trigger macros and commands via natural language requests to their connected AI tools after enabling the feature in app preferences.
TIIUAE has released Falcon Perception, a 0.6B-parameter early-fusion Transformer that combines image patches and text in a single sequence for open-vocabulary object grounding and segmentation. The model achieves 68.0 Macro-F1 on SA-Co (vs. 62.3 for SAM 3) and introduces PBench, a diagnostic benchmark that isolates performance across five capability levels. TIIUAE also released Falcon OCR, a 0.3B model reaching 80.3 on olmOCR and 88.6 on OmniDocBench.