🚀 RELEASES & LAUNCHES
Meta releases Muse Spark Open, an 816B-parameter open-weight model that runs on Mac Studios
Meta released Muse Spark Open, an open-weight multimodal mixture-of-experts model it calls the most capable US open-weight model, optimized with Apple to run on Mac Studios via MLX.- Aimed at local/developer users: Meta released Muse Spark Open on October 15, 2026, claiming it is the most capable US open-weight model to date.
- Specs: 816B total parameters with 42B active per token, a 1,048,576-token context window shared across text, image, audio, and video inputs; trained for agentic coding like multi-file editing, debugging, and tool use.
- Meta says it is the first model in its class to run on a desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework, the result of a Meta-Apple collaboration.
- Ships with custom Helion GPU kernels in SGLang (1.03-3.95x kernel speedups on NVIDIA and AMD GPUs), and runs on SGLang, vLLM, TensorRT-LLM, and Hugging Face Transformers (added in v5.20.0); checkpoints available in BF16 and NVFP4.
- It matters because it brings frontier-class agentic and multimodal capabilities to open weights and local desktop inference, following Meta's closed Muse Spark 1.1 API launch in July 2026.
llama.cpp v0.7.0 adds LiliCorr speculative decoding, MoE host-memory cache, and new models
llama.cpp released v0.7.0, adding LiliCorr speculative decoding, a GPU cache for host-memory MoE experts, and support for a batch of new models.- Targets local LLM inference users across CPU, CUDA, Metal, Vulkan, and other backends, plus apps like Jan and Llama that embed llama.cpp as their engine.
- Adds the LiliCorr speculative decoding technique and MTP support for GLM5-Next.
- Introduces a new GPU cache for MoE experts kept in host memory via --moe-cache-mib, delivering up to ~2.2x decode speedup when experts are offloaded.
- Adds support for K2 Horizon (dense and MoVA), Prism Bonsai 2 27B, EmbeddingGemma-2 (text+vision+audio), LiquidAI d1/d1-omni, MiniCPM-V 4.7, Nemotron Super 3.5 VL, and more.
- Server gets a refactored chat session API, better slot and prompt-cache restore, a new default port (9931), a web UI models manager, and ggml is updated to v0.27.0.
Sources 📎 Coverage: github.com · 👁 via llama.cpp releasesNVIDIA releases open Nemotron 3.5 Super VL 120B-A12B vision-language model for agents
NVIDIA released Nemotron 3.5 Super VL, an open 120B-parameter (12B active) vision-language MoE model for agentic workflows, alongside edge-focused Nemotron 3.5 Bolt models.- Aimed at developers and enterprises building always-on agents for document intelligence, coding, search, cybersecurity, and finance workflows.
- Super VL is a 120B-parameter MoE with 12B active parameters, a 1M-token context window, image/video input, tool calling, and toggleable reasoning; trained on ~26 trillion tokens with ~3.8M post-training samples.
- NVIDIA claims it is the fastest open agentic model in its class at up to 5x the throughput of peers, with accuracy comparable to its largest reasoning model at roughly one-fifth the size; it scores 1190 Elo on GDPval vs 493 for Nemotron 3 Super.
- Released in BF16 and NVFP4 on Hugging Face under the openmdw-1.1 license with open datasets and recipes; available on Baseten, Kilo, and via NIM, with Bolt 74B-A9B 24GB out now and Bolt E4B 4GB coming Oct. 23.
- Dozens of organizations including Scale AI, AI21, Atlassian, CrowdStrike, Perplexity, and Salesforce are post-training or integrating it, giving adopters an open, self-hostable multimodal agent model.
Black Forest Labs releases FLUX.3 Image Klein, a 5.6B open-weights image model
Black Forest Labs published open weights for FLUX.3 Image Klein, a 5.6B text-to-image and image-editing model, on Hugging Face.- Aimed at local/developer users: one model handles text-to-image, single-reference edits, and multi-reference compositions from up to 10 reference images.
- 5.6B-parameter open-weights model using Qwen3-VL-4B-Instruct as its text encoder; weights live on Hugging Face under black-forest-labs/flux-3-image-klein.
- It is the klein (small) variant of the FLUX 3 multimodal family, following the earlier FLUX.2 klein models that ran sub-second on ~13GB VRAM consumer GPUs.
- Open-weight klein releases give builders a compact, fine-tunable image model from the FLUX 3 generation without needing API access.
Mistral releases Timestral, its first family of time-series forecasting models (5M–209M)
Mistral released Timestral, a family of four open-weight time-series foundation models for zero-shot probabilistic forecasting that run on CPU.- For developers and enterprises doing forecasting on sensor telemetry, observability metrics, energy load, and demand data.
- Four sizes shipped: 5M, 28M, 123M, and 209M parameters, under a Modified MIT licence, in public preview on Hugging Face and Mistral docs.
- Supports univariate and multivariate forecasting, past-only and known-future covariates, categorical covariates, and probabilistic outputs across 21 quantiles.
- Accepts up to 32,768 historical time steps and forecasts up to 1,024 steps in one forward pass; CPU-only inference takes 10 ms (5M) to 140 ms (209M) for a day-ahead forecast.
- Mistral claims the 209M model ranks first among pretrained zero-shot models on GIFT-Eval, TIME, and fev-bench, beating TimesFM-3 and Falcon-2.0 with fewer parameters; a technical report is to follow.
Bluesky expands AI app Attie to let anyone vibe-code websites, games, and apps
Bluesky expanded its AI-powered Attie app so users can build interactive experiences like websites, games, and quizzes on the social network.- Aimed at Bluesky's 40 million registered users and developers on the open AT Protocol network, building on Attie's original feed vibe-coding.
- The expanded Attie lets anyone create interactive experiences — websites, games, quizzes, and mini-apps — described in plain English, powered by Anthropic's Claude.
- Attie launched at the ATmosphere 2026 conference in closed beta and faced heavy backlash, becoming Bluesky's second most-blocked account with 125,000 blocks versus 1,500 followers.
- Bluesky positioned Attie as user-controlled AI on an open protocol, contrasting with opaque ad-driven algorithms on X and Instagram.
- The expansion marks Bluesky's push toward full social mini-apps built by non-coders on top of its decentralized network.
Sources 📎 Coverage: techcrunch.com · 👁 via TechCrunch AIgoogle-universal-cart — TechCrunch AI
🔬 RESEARCH
DeepMind details antimicrobial peptide discovery using Google Antigravity
Google DeepMind published a walkthrough showing biologists used Google Antigravity to identify a novel antimicrobial peptide candidate now undergoing wet-lab validation.- Aimed at biologists and drug-discovery researchers fighting antimicrobial resistance, projected to cause 10 million annual deaths globally by 2050.
- Google Antigravity served as an agentic science workbench (with Gemini 3.8 Flash), compressing cross-database search, structural analysis, and literature synthesis into one session.
- The agent used ProtNLM, a model predicting protein function from amino acid sequences, to annotate millions of previously unnamed UniProt entries.
- It identified a 97-amino-acid bacteriocin (UniProt ID A0A0U0D591) in Streptococcus pseudopneumoniae with predicted activity against respiratory-infection pathogens, verified via BLAST, AlphaFold/Foldseek structural checks, and flanking-gene analysis.
- The lead is now undergoing wet-lab validation, demonstrating AI agents can produce biologically plausible therapeutic candidates.
- The lead is now undergoing wet-lab validation, demonstrating AI agents can produce biologically plausible therapeutic candidates.
Microsoft Research unveils Magnetar, a post-quantum zkVM proving RV64 programs from standard hashes
Microsoft Research released Magnetar, a transparent zkVM that proves RV64I program executions with hash-based post-quantum security and major speed gains.- Aimed at developers proving general-purpose program executions, including a zkEVM for existing Ethereum smart contracts, with quantum-resistant security.
- Magnetar is the first zkVM combining general-purpose RV64I proving, security rooted only in standard hashes like SHA-256/SHA-3, and efficient recursion, without elliptic-curve or lattice assumptions or circuit-friendly hashes like Poseidon.
- At 120 bits of provable security on a 12-core M4 Max MacBook Pro, it proves 2^22-cycle executions at 7.2–8.7 million cycles per second, with ~600 KiB proofs verifying in 11–13 ms.
- Across 15 workloads it outperforms SP1 Hypercube by 24–66×, Jolt-Dory by 12–37×, and Jolt-Akita by 6–20× (geometric means ~39×, 18×, 9.6×); an optimized version reaches 9.5–12.8 million cycles per second.
- Implemented in Rust by Srinath Setty, the paper is on the Cryptology ePrint Archive (2026/2423), reducing reliance on non-standard cryptographic assumptions for verifiable computing.
- Related Microsoft Research work announced alongside it includes Vega, a zero-knowledge credential system proving facts like age from mobile driver's licenses in 92 ms with no trusted setup.
📄 Source for Anthropic tells investors Q3 revenue hit $16.8bn, up nearly 50%, ahead of IPO — anthropic.com
📄 Source for NVIDIA releases Nemotron 3.5 Bolt 74B-A9B 24GB for local agentic AI — huggingface.co
📊 *