Model Releases

Meta Muse-Glimmer-30B: 30B Parameter Multimodal Model with Apache 2.0 License — Meta drops a 30B image-text-to-text conversational model under Apache 2.0. Two arXiv papers document the architecture (2504.13181, 2602.06036), plus transformers and safetensors support out of the box. Endpoints compatibility means you can deploy this without wrestling custom serving code. 🤖

Open Source Releases

cline/cline v4.1.8: Manual Vertex model ID entry, Fable 5 support, full region catalog — Finally lets you paste any Vertex model ID manually, including unlisted ones. Adds Fable 5 support and shows the full model catalog per Vertex region instead of a hardcoded global list, with region validation baked in. Small quality-of-life wins that add up.

alpiecode 0.9.2 - Autonomous AI Coding Agent CLI with Local VLM — CLI coding agent powered by a local 169Pi GGUF vision-language model. Fully offline, no cloud deps. If you’ve been itching for autonomous code gen/editing without sending your codebase to an API, this is worth a spin. 🛠️

Research Worth Reading

Training Variable Long Sequences with Data-Centric Parallel — Data-centric parallelism for variable-length long sequences. Tackles workload imbalance in static configs without the headache of dynamic methods. Practical path to efficient long-context training without rewriting your whole stack. 📄

The Knowing-Saying Gap: When Probes See Errors that Confidence Misses — Linear probes detect corrupted context with near-perfect accuracy, yet can’t predict final answer correctness. A fundamental dissociation with real implications for monitoring deployed LLMs. Your confidence scores might be lying to you.

Dynamic Coalition Formation and Communication Pricing in Skill-Based Agentic AI Systems — Models agent coalition formation and communication pricing to optimize token cost, latency, and error propagation in multi-agent systems. Framework for efficient communication topologies in heterogeneous skill-based architectures. Multi-agent folks, take notes.

CODS: Iterative Bellman-Residual Data Selection for Reusable Offline Reinforcement Learning — Critic-guided data selector alternating between fitting an algorithm-matched critic and selecting high Bellman-residual transitions. Cuts redundant computation across seeds/hyperparams while preserving rare transitions critical for long-horizon credit assignment. Offline RL practitioners, this one’s for you.

Shape Mutating Expert Compression: LorExperts and BTExperts — Two MoE compression methods: low-rank expert decomposition (LorExperts) and block-term tensor decomposition (BTExperts). Preserves all experts, avoids router retraining, beats pruning/merging on accuracy-compression trade-offs. MoE compression that doesn’t require a full retrain cycle. 🔥

LUCID: Latent-Skill Unified Control via Imagined Dynamics for Long-Horizon Humanoid Loco-Manipulation — Unified control using latent skills and imagined dynamics for long-horizon humanoid loco-manipulation. Ditches scripted planners and task-specific policies for flexible skill composition and high-level decision making. Robotics + latent skills = actually generalizable policies.

AI Dev Tools

stablyai/orca: ADE for parallel agent fleets with multi-platform support — Agent Development Environment for running fleets of parallel coding agents on your own subscriptions. Desktop, mobile, VPS deployments with worktree-based isolation and terminal-native UX. If you’re orchestrating multiple agents, this handles the infrastructure grunt work. 🛠️

anthropics/claude-code v2.1.225: Gateway spend-limit support and workspace trust prompt for agents — Adds gateway spend-limit warnings with cap name, reset time, and operator message. Workspace trust prompt for claude agents in untrusted directories. Fixes transient 401 replacement bug. Guardrails for when you let agents loose in your codebase.

cline/cline v4.1.7: Restored View Changes button, copy button, pre-registered OAuth for MCP — Restores ‘View Changes’ button on completion rows via SDK checkpoints. Copy button on turn-final responses. Pre-registered OAuth clients for remote MCP servers where dynamic registration isn’t an option. Steady iteration on the DX.

Today’s Synthesis

Meta’s Muse-Glimmer-30B drops a 30B Apache 2.0 multimodal model with endpoints-ready serving — pair that with alpiecode’s local VLM agent pattern (169Pi GGUF, fully offline) and you’ve got the building blocks for a private, multimodal coding agent that never leaves your network. Now scale it: stablyai/orca gives you worktree-isolated, terminal-native fleet orchestration across desktop, mobile, or VPS. The missing piece? Compression that doesn’t require retraining. LorExperts/BTExperts deliver exactly that — low-rank and block-term decomposition preserving all experts and router weights, beating pruning on accuracy-compression trade-offs. Compress Muse-Glimmer or your local VLM of choice, spin up an orca fleet, and you’re running a private, multimodal agent swarm for the cost of compute — no API keys, no data egress, no spend-limit surprises. 🤖🛠️🔥