Sunday, April 12, 2026

TLDR

Meta launches Muse Spark (Apr 8) — first model from Meta Superintelligence Labs under Alexandr Wang; multimodal reasoning, rolling out across WhatsApp/Instagram/FB
Z.ai's GLM-5.1 tops SWE-BenchA benchmark that tests whether a model can resolve real GitHub issues by editing code in popular Python repositories.Jimenez et al. 2023 ↗ Pro (Apr 7) — open-source (MIT) model beats Claude Opus 4.6 and GPT-5.4 on the hardest coding benchmark
Google ships Gemma 4 (Apr 2) — open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ multimodal models (Apache 2.0) with agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ workflow support, runs on-device
Flowise CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗ 10.0 RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ actively exploited (Apr 8) — CVE-2025-59528; 12,000+ instances exposed; patch to v3.0.6 immediately
Claude 3 Haiku retires Apr 19 — migrate to Haiku 4.5 this week

Action Items

Patch Flowise to v3.0.6 immediately if used anywhere — CVE-2025-59528 (CVSS 10.0) is under active exploitation with 12K+ exposed instances (The Hacker News)
Migrate off Claude 3 Haiku before Apr 19 — retirement date is one week away; switch to Claude Haiku 4.5 (Anthropic Docs)
Migrate off Gemini 2.0 Flash before June 1 — final deprecation of Gemini 2.0 Flash and Flash-Lite; move to Gemini 2.5 Flash or 3.1 Flash-Lite (FindSkill.ai)
Review Gemini API billing if on free tier — Pro models now paywalled; free tier restricted to Flash models only since Apr 1 (LaoZhang AI)

Models & Tools

Meta Muse Spark (Apr 8) — Meta Superintelligence Labs' first model: natively multimodal reasoning with visual chain-of-thought, tool use, and multi-agent orchestration. Small and fast by design, it ranks 4th on the Artificial Analysis Intelligence Index behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6. Now live on meta.ai; rolling out to WhatsApp, Instagram, and FB. (TechCrunch)
Z.ai GLM-5.1 (Apr 7) [Open Source] — MIT-licensed model that scored 58.4 on SWE-Bench Pro, topping Claude Opus 4.6 (57.3) and GPT-5.4 (57.7). Can work autonomously on a single coding task for 8+ hours. First Chinese open-source model to lead global coding benchmarks. (TechBriefly)
Google Gemma 4 (Apr 2) [Open Source] — Apache 2.0 multimodal family (27B dense, 26B MoEMixture of expertsA model made of many sub-networks, called experts, where a learned router sends each input to only a few of them.Shazeer et al. 2017 ↗, edge E2B/E4B variants). Handles text, images, and audio natively with agentic workflow support. New AICore Developer Preview for Android and litert-lm CLI for desktop/Raspberry Pi. (Google Blog)
Google TurboQuant (Apr ~8) [Infra] — Presented at ICLR 2026; significantly reduces KV cache memory overhead, enabling models with massive context windowsHow much text a model can see and use at once while writing a reply.Anthropic ↗ to run far more efficiently. Directly relevant if you're self-hosting large models. (Google Developers Blog)
Meta plans open-source Avocado & Mango (Apr 6) — Axios reports Meta is building two proprietary frontier models (Avocado LLMLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗, Mango multimedia generator) with open-source variants to follow. Largest models will stay proprietary. (Axios)

APIs & Pricing

Claude 3 Haiku retirement: Apr 19 — Legacy model ($0.25/$1.25 per MTok) goes offline. Migrate to Haiku 4.5. (Anthropic Docs)
Gemini 2.0 Flash/Flash-Lite: final deprecation June 1 — Migrate to Gemini 2.5 Flash or 3.1 Flash-Lite. (FindSkill.ai)
Gemini API free tier restricted (Apr 1) — Pro models now require paid API key; free tier limited to Flash models only. Mandatory spend caps enforced: Tier 1 $250/mo, Tier 2 $2K/mo. (LaoZhang AI)
Anthropic Opus pricing dropped 67% (recent) — Opus context window expanded to 1M tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗. Industry-wide API costs down 40-70% across major providers. (AI Empire Media)

Security

Flowise CVE-2025-59528 (Apr 8) [Trending] — CVSS 10.0 RCE in the CustomMCP node; executes user-provided JavaScript without validation, granting full Node.js runtime access. Active exploitation detected from Starlink IPs. 12,000-15,000 instances exposed. Patch: upgrade to v3.0.6. This is Flowise's third in-the-wild exploit. (Bleeping Computer)
Prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ surges 340% in 2026 (Apr 9) — OWASP LLM Security Project classifies prompt injection as the single highest-severity vulnerability for deployed LLMs. CIS highlights indirect prompt injection via documents/emails as a growing risk in government GenAI deployments. (Help Net Security)

Industry & Business

Q1 2026 venture funding: $300B (Apr ~7) [Trending] — Record-shattering quarter; AI accounted for $242B (80%). Four of the five largest VC rounds ever: OpenAI ($122B), Anthropic ($30B), xAI ($20B), Waymo ($16B). (Crunchbase)
Eclipse raises $1.3B (Apr 7) — Cerebras backer Eclipse raised $1.3B for robotics and AI infrastructure investments. (Bloomberg)
Meta AI capex: $115-135B for 2026 (Apr 8) — Nearly 2x last year, reflecting massive infrastructure buildout for Muse Spark and future models. (CNBC)
Tech layoffs: 78,500 in Q1 (Apr ~10) [Hiring] — 48% attributed to AI/automation. Oracle (20-30K), Amazon (16K), Dell (~11K) lead cuts. Meanwhile, AI/ML job postings up 163% YoY; Databricks has 841 open roles. (Tom's Hardware)

AI in Production

78% of enterprises have AI agent pilots, <15% reach production (Mar 2026) [Agents] — Five gaps account for 89% of scaling failures: integration complexity, inconsistent output quality, missing monitoring tooling, unclear ownership, and insufficient domain training data. (Machine Learning Mastery)
OpenAI hits $25B ARR, enterprise at 40% (Apr ~10) — OpenAI Frontier helping Oracle, State Farm, and Uber deploy agents company-wide; consumer-enterprise parity expected by EOY. (OpenAI)

Policy & Safety

Colorado AI Act takes effect June 30 — SB 24-205 requires developers of high-risk AI systems to provide model cards, dataset documentation, and impact assessments. 90-day disclosure window for discovered risks. NIST AI RMF or ISO/IEC 42001 compliance serves as affirmative defense. (Brownstein)
California AI procurement EO (Apr ~7) — Governor Newsom signed EO N-5-26 directing state agencies to develop standards for AI vendors seeking government contracts, plus expanding responsible GenAI use in state operations. (DLA Piper)
600+ state AI bills introduced in 2026 sessions — DOJ's new AI Litigation Task Force established to challenge state AI laws that unconstitutionally regulate interstate commerce. EU AI Act rules expected in August. (Wilson Sonsini)

Paper Spotlight

Neuro-Symbolic AI Cuts Energy 100x While Boosting Accuracy — Tufts researchers combined neural networks with symbolic reasoning for robot control, achieving 95% task success (vs 34% baseline) while using just 1% of training energy and 5% of inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ energy. Presenting at ICRA Vienna in May. (ScienceDaily)

openclaw/openclaw 247K+ stars — Open-source personal AI assistant that runs locally and connects to 50+ messaging platforms (WhatsApp, Slack, Discord, iMessage, etc.). The viral breakout project of 2026. (GitHub)

AI news, Sunday, April 12, 2026