AI Daily Dev — April 30, 2026

Nvidia Open-Sources Nemotron 3 Nano Omni, a 30B Multimodal MoE With 9x Throughput

Hybrid Mamba-Transformer MoE: 30B total, 3B active, 256K context, vision/audio/video/text in a single inference loop.
Tops six open-model leaderboards — MMlongbench-Doc, OCRBenchV2, WorldSense, DailyOmni, VoiceBench, MediaPerf — at ~9× the throughput of peer omni models.
BF16, FP8, NVFP4, and Unsloth GGUFs on Hugging Face day one; community report of 1M-context 25 t/s on a $400 RX 6700 XT.
Free tier on OpenRouter; Foxconn, Palantir, Oracle, and H Company already in production; Crusoe and Clarifai serving at 400 t/s zero-day.

open-source nvidia.com

33B-total / 3B-active MoE, fully in-house training stack, 30T tokens, sized to fit a 36 GB Mac via Ollama.
44.5% SWE-bench Pro, 30.1% Terminal-Bench 2.0, 68.2% SWE-bench Verified — beats Claude Haiku 4.5 (39.5%) and dense Gemma 4 31B (35.7%) on Pro.
Launch-day support in vLLM, Transformers, and NVIDIA TRT-LLM; weights on Hugging Face, free on OpenRouter, in Ollama and Puter.js.
Companion proprietary Laguna M.1 (72.5% SWE-bench Verified) free via API for a limited time; HN front page item 47936511.

models poolside.ai

What's Next with AWS 2026 in San Francisco, April 28 — first concrete shipping of the post-Microsoft OpenAI distribution rights.
GPT-5.5 and GPT-5.4 in preview on Amazon Bedrock; Codex on Bedrock authenticates with AWS creds and counts toward AWS commitments.
Bedrock Managed Agents (limited preview) is built on the OpenAI harness — first non-Microsoft cloud-native OpenAI agent service.
Amazon Quick relaunches as a local-first desktop app: free tier with no AWS account, $20/user/mo Plus; native Google Workspace, M365, Slack, Salesforce.

industry aws.amazon.com

WSJ scoop, picked up by Bloomberg April 30: Trump admin officials told Anthropic they don't agree with expanding Mythos from ~50 to ~120 entities.
Two stated objections: cyber-offensive capability proliferation, and Anthropic compute capacity that would dilute government priority access.
Lands two days after the Anthropic and OpenAI classified briefings to House Homeland Security on cyber-capable models.
Bloomberg Opinion frames the public spat as already 'eroding national security efforts' that depend on government use of Mythos.

industry bloomberg.com

Durable, observable orchestration layer baked into Mistral Studio and Le Chat; Python workflows publish to Le Chat for org-wide use.
Built on Temporal with AI-specific extensions: streaming, large payload handling, traceable execution, human-in-the-loop approvals.
Mistral says it's already running millions of executions per day across ASML, ABANCA, CMA-CGM, France Travail, La Banque Postale, and Moeve.
Orchestrator is Mistral-managed, but execution workers and data stay in customer cloud, on-prem, or hybrid environments.

frameworks mistral.ai

Two new groups: Agentic Authentication (delegation + phishing-resistant auth) and Payments (agent-initiated commerce, chaired by Mastercard and Visa).
Google donates Agent Payments Protocol (AP2) v0.2 — adds 'Human Not Present' transactions — as the seed spec for the payments group.
Mastercard and Google co-donate Verifiable Intent, a tamper-evident log of user-authorized agent actions, into the same workstream.
First serious industry move to standardize agent identity and accountability the way FIDO standardized passkeys for humans.

industry fidoalliance.org