Agentic AI News Oct. 16, 2026

Muse Spark Open. Meta released Muse Spark Open, an 816B-parameter open-weight Mixture-of-Experts model with 42B active parameters, 1M-token context, and text/image/video/audio input, built for coding and long-horizon agentic work. It runs on two Apple M5 Ultra Mac Studios via MLX or on 4 Blackwell Ultra/8 H200s with NVFP4, scores 89.2 on MCP-Atlas, and is supported by Unsloth, SGLang, vLLM, llama.cpp, and Apple MLX.

• Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context(opens in a new tab)
• Meta launches an open-weight AI model that runs on Macs(opens in a new tab)
• Muse Spark Open is Meta's answer to Qwen, GLM, and Kimi in the open-weight race(opens in a new tab)
• Meta and Apple team up to make Muse AI models run better on Macs(opens in a new tab)
• Introducing Muse Spark Open: the most capable US open-weight model to date(opens in a new tab)
• Unsloth Muse Spark Open is out!(opens in a new tab)
• UNSLOTH MUSE SPARK OPEN!!(opens in a new tab)

Nvidia Nemotron 3.5 Super VL. Nvidia released Nemotron 3.5 Super VL, a 120B MoE vision-language model activating 12B parameters per token for always-on enterprise agents, with 1M context and NVFP4/BF16 checkpoints. Nvidia also shipped two edge Bolt variants (74B-A9B compressed to 24GB and a 4GB version) and a Nemotron-3.5-Embed-1B multilingual/multimodal retrieval model, plus open datasets and training recipes.

• NVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model(opens in a new tab)
• Nvidia's Nemotron 3.5 Super VL is an open model built to run enterprise AI agents nonstop(opens in a new tab)
• nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B model family released (base, NVFP4, BF16, and LiLiCorr speculative decoding)(opens in a new tab)
• nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB · Hugging Face(opens in a new tab)
• nvidia/Nemotron-3.5-Embed-1B-BF16 · Hugging Face(opens in a new tab)

IBM Granite Switch. IBM released Granite Switch 4.1 in 8B and 30B variants, composing a base model with Activated LoRAs for chat, decision-making, guardian safety, and speech transcription in one endpoint. The approach keeps VRAM near that of the base model and reuses KV cache across tasks.

• Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint(opens in a new tab)
• ibm-granite/granite-switch-4.1 30B/8B(opens in a new tab)

Ternary Bonsai 2. Prism ML released Ternary Bonsai 2 models in 2B, 4B, and 9B sizes based on Qwen 3.5, with dFlash/MTP support and a WebGPU demo running the 2B entirely in-browser at up to 500 tok/s on an M4 Max.

• Ternary Bonsai 2 Family: 2B, 4B, & 9B based on Qwen 3.5(opens in a new tab)
• Ternary Bonsai 2 (2B, 4B, and 9B) just released on Hugging Face. The 2B can run entirely locally in the browser at up to 500 tok/s on an M4 Max with custom WebGPU kernels.(opens in a new tab)

FLUX 3 Image Klein. Black Forest Labs released FLUX 3 Image Klein, a 5.6B open-weights image model for text-to-image, image editing, and generation with up to 10 reference images. It supports native 4K output, pixel-accurate editing, and an INT8 build that runs on 8GB VRAM in under 9 seconds on an RTX 4070.

• Flux.3 Image Klein is now out(opens in a new tab)
• Black Forest Labs' open-weights image model FLUX 3 Image Klein runs on consumer GPUs with just 8 GB of VRAM(opens in a new tab)

Mistral Timestral. Mistral released Timestral, a family of time-series foundation models from 5M to 209M parameters for zero-shot forecasting, accepting up to 32,768 historical steps and forecasting up to 1,024 steps with probabilistic quantile outputs.

• mistralai/Timestral(opens in a new tab)

Google Gemini 4 Argon. Google introduced Gemini 4 Argon as a single model for robotics that controls robots, generates interactive 3D training environments, and evaluates training data. Demos still require detailed human prompting and are available through paid developer access programs.

• Google wants Gemini 4 Argon to be the only AI model robots need(opens in a new tab)

LangChain LangSmith updates. LangChain shipped LangSmith AI Search for semantic trace queries, Managed Deep Agents v0.9 with Slack reactivity and scheduling, Engine v2 red teaming, and Trajectories for chronological agent session views.

• October 2026 LangChain Newsletter: Everything we shipped at Interrupt NYC & London(opens in a new tab)

CopilotKit CLiF. CopilotKit released CLiF in early access, a continual learning platform that captures full-stack agent and user trajectories and turns them into skills, production evals, and fine-tuning datasets.

• CopilotKit Releases CLiF: A Continual Learning Platform That Turns Agent Traces and User Actions into Skills, Evals, and Fine-Tuning Data(opens in a new tab)

BlenderWorld benchmark. Alibaba's Qwen team published BlenderWorld, a benchmark measuring how well visual agents create 3D scenes in Blender from real-world references, signaling possible Qwen focus on agentic 3D design via MCP.

• BlenderWorld Benchmark from Alibaba(opens in a new tab)

DataInventor. DataInventor, an open-source dataset generator on Hugging Face Spaces, added a Visualize mode for inspecting generated data.

• DataInventor: open-source dataset generator on HF Spaces, now with a Visualize mode(opens in a new tab)

Drama3 in audio.cpp. Drama3 support landed in audio.cpp via a community contribution, reaching up to 5× real-time on an RTX 5090 with CUDA optimizations. Vulkan and Metal support exist but still need iteration for stability and correctness.

• Drama3 in audio.cpp: Up to 5× faster than real time at full speed, or 2× real-time speed with < 8 GB of VRAM on an RTX 5090!(opens in a new tab)

Interpretability via agents. Goodfire's Nick Cammarata argues that neural networks are already clean: many polysemanticity and superposition cases dissolve under further agentic analysis, suggesting interpretability may scale without more bitter lessons and lowering P(doom).

• [AINews] Agents disprove Polysemanticity; greatly improve Interpretability & Alignment(opens in a new tab)

Diffusion language model. ALoDLM 8b, a looped diffusion language model based on Qwen3-8B, shows about 3× inference speed over its base model with better benchmark results.

• ALoDLM 8b looped diffusion language model: ~3x inference speed vs its base model qwen3-8b with better benchmark results(opens in a new tab)

Indic pretraining data. NVIDIA, IIT Madras, and AI4Bharat introduced Nemotron-pretraining-Indic-v1, a roughly 2.8 trillion token pretraining corpus across 22 Indian languages, with an open GPU-accelerated data generation pipeline.

• Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline(opens in a new tab)

Physics benchmark GPT-6-Astra. TheoPhysEval benchmark found GPT-6-Astra solved 49 of 50 theoretical physics research problems at under $3 each, matching leading specialists; researchers note a missing sense of 'taste' limits breakthroughs.

• GPT-6-Astra solves 49 of 50 physics research problems at $2.40 apiece(opens in a new tab)

Antimicrobial peptides. Google Antigravity, an agentic science workbench, helped biologists identify a novel antimicrobial peptide candidate with predicted activity against respiratory pathogens, now in wet-lab validation.

• Accelerating the search for antimicrobial peptides with Google Antigravity(opens in a new tab)

Transformer as game console. A 32M-parameter decoder-only transformer, Chipformer, implements a CHIP-8 game console via compiled circuits, running Pong, Tetris, Space Invaders, and 116 other games through next-token prediction without an emulator.

• I turned a transformer into a game console that runs Pong, Tetris, Space Invaders and 116 more games.(opens in a new tab)

DeepSeek on four Radeons. DeepSeek-V4-Flash-0731 runs on four AMD Radeon AI PRO R9700 cards at about 79 tok/s using the affinity engine with all experts resident, after a prior two-card limit was bypassed on real hardware.

• **DeepSeek-V4-Flash-0731 on 4x Radeon AI PRO R9700 with affinity: the "refused" 4-card config works, ~79 tok/s on real cards**(opens in a new tab)

Snapdragon mini PC. A £300 Lenovo IdeaCentre Mini with Snapdragon X1-26-100 runs a Qwen3.6 35B coder at about 24.5 tok/s (30 with MTP) and Qwen3-Next 80B at 17-18 tok/s via CPU-only llama.cpp and ARM NEON kernels.

• What a £300, 35W Snapdragon mini PC can actually do 👀: 35B coder at ~30 tok/s, and an 80B model that doesnt even fit in RAM running at 17 tok/s(opens in a new tab)

Qwen on two GPUs. capy-flash, a tensor-parallel engine, runs Qwen3.8-Flash-Next 125B-A6B in full NVFP4 on two 16GB RTX 5070 Ti GPUs with experts offloaded to RAM, reaching roughly 100 tok/s generation.

• capy-flash: Qwen3.8-Flash-Next (125B MoE) in full NVFP4 on two 16 GB consumer GPUs(opens in a new tab)
• FP4 QAD quant of Qwen3.8-Flash-Next from Nvidia(opens in a new tab)

RTX Spark laptops. HP's RTX Spark laptop with 128GB RAM is available for $3,832 after coupon, and Microsoft is optimizing Windows for RTX Spark's Arm CPU, Blackwell GPU, and unified memory for local AI agents and gaming.

• 128GB RTX Spark laptop for $3832.(opens in a new tab)
• How Windows is being optimized for Nvidia’s RTX Spark laptops(opens in a new tab)

ChatGPT Chrome tabs. ChatGPT's Chrome extension can now view open browser tabs when tagged, letting users ask for comparisons, form filling, or summaries across tabs from the web or desktop app.

• ChatGPT can peek at your Chrome tabs now(opens in a new tab)

Bluesky Attie. Bluesky expanded its Attie AI tool to let users generate small social games, polls, dashboards, and websites from prompts, shareable on attie.site domains and built on AT Protocol.

• Bluesky&#8217;s AI tool now lets you make games and apps to share on social(opens in a new tab)
• Bluesky expands its AI-powered Attie app to let anyone build websites, games, and more(opens in a new tab)

Google Universal Cart. Google launched Universal Cart in the US, letting users save products from Search and Gemini, get price-drop and back-in-stock alerts, and use credit-card perks and promo codes at checkout through Universal Commerce Protocol.

• Google’s Universal Cart shopping tool launches in the US(opens in a new tab)

YouTube Music AI labels. YouTube Music is rolling out generative AI labels for songs created entirely or altered significantly with AI, similar to labels on AI-generated videos on YouTube.

• Singer or clanker? YouTube Music will tell you with new gen AI labeling.(opens in a new tab)

Rivian AI agents. Rivian plans to let third-party AI agents from companies like Anthropic run in its vehicles, building on its in-house agentic framework that routes voice commands to car software.

• AI agents are coming to Rivian’s cars(opens in a new tab)

Amazon ambient AI. Amazon devices chief Panos Panay says ambient AI wearables like the new Bee devices will shift computing from app-driven 'point, shoot, return' to background agents, with Alexa and Leo integration.

• Amazon devices chief Panos Panay sees a future where AI devices replace the need to use apps(opens in a new tab)

AI oversight steps up. The FTC sent civil investigative demands to OpenAI and Anthropic over recent agent safety incidents, while Anthropic granted METR researchers employee-level access for independent risk assessment. Bank of America has 4,000 people fixing vulnerabilities uncovered by Anthropic's Mythos and other AI models.

• Report: FTC probes Anthropic, OpenAI after Trump dismissed AI fears as “hoax”(opens in a new tab)
• Anthropic is letting METR researchers look under the hood with employee-level access(opens in a new tab)
• Bank of America puts 4,000 people on fixing vulnerabilities uncovered by Anthropic's Mythos and other AI models(opens in a new tab)

AI trust backlash. New films portray Big Tech founders critically, a Trump post included an AI-edited White House photo, and celebrity deepfakes continue spreading. Concerns are growing over agents like Meta's Muse pulling in friends' data without consent, while Flock's ALPR surveillance faces scrutiny from lawmakers and investigators.

• The tech backlash is now in theaters(opens in a new tab)
• Trump posts AI-edited photo to brag about White House restoration work(opens in a new tab)
• Nobody is safe from celebrity AI deepfakes(opens in a new tab)
• Friends don&#8217;t let friends feed their data to AI agents(opens in a new tab)
• Benn Jordan on Flock, privacy, and the new surveillance backlash(opens in a new tab)

AI upscaling flop. Sega pulled Total War: Shogun 2 Complete Edition from storefronts after players found bugs and AI-upscaled artwork, showing the quality risks of AI in game remasters.

• Sega pulls Total War: Shogun 2 remaster from sale following disastrous launch(opens in a new tab)

Cal AI influencer growth. Cal AI generated over $30 million in two years and was acquired by MyFitnessPal using a small army of paid influencers; founder Zach Yadegari says every post drove roughly linear new user growth.

• When it came to Cal AI’s viral success, a small army of influencers paid off(opens in a new tab)

GPT-Rosalind academic plan. OpenAI launched the Rosalind Academic Plan, offering eligible labs up to 30 Business Standard seats at no cost for up to 12 months, with sponsored access to GPT-Rosalind for life sciences research.

• Bringing GPT-Rosalind to academic research labs(opens in a new tab)