Model Releases

  • nvidia/LocateAnything-3B — NVIDIA’s 3B-parameter vision-language model fine-tuned from Qwen2.5-3B-Instruct for object detection and visual grounding. Custom code support, multiple arXiv papers backing it — worth a look if you’re building Eagle-vision or grounding pipelines.

  • LiquidAI/LFM2.5-8B-A1B — 8B MoE model from LiquidAI targeting edge deployment with 7-language support. Eval results included, TIG-compatible, backed by arXiv 2511.23404. If you’re shipping models to resource-constrained environments, this one’s on the radar. 🤖

Open Source Releases

  • MOSS-TTS — Open-source speech and sound generation model family from MOSI.AI and the OpenMOSS team. Targets long-form speech, multi-speaker dialogue, voice design, environmental sound effects, and real-time streaming TTS. The “high-fidelity, high-expressiveness” pitch is bold — but the feature set is legitimately broad.

  • diffcb 0.1.4 — Silverman’s modality test as a differentiable PyTorch layer with IFT backward pass. Making a classical statistical test end-to-end differentiable is the kind of niche-but-clever contribution that makes you go “oh, I could actually use that.” 📄

  • Crawl4AI — Open-source web crawler and scraper designed to be LLM-friendly. Streamlines feeding web data into LLM pipelines without the usual scraping duct-tape. If your RAG pipeline is held together by BeautifulSoup and prayers, this might help.

  • OpenCode v1.15.12 — Adds ACP integrations for sending prompts, slash commands, and usage updates via acp-next, plus experimental WebSocket transport for OpenAI responses. Adaptive reasoning controls for Anthropic Opus 4.7+ models round it out.

  • humecodec 0.7.0 — FFmpeg integration for PyTorch with bundled libraries. Native PyTorch tensor handling of multimedia streams without manual FFmpeg subprocess juggling. For audio/video ML engineers, this closes a surprisingly common gap. 🛠️

Research Worth Reading

  • Behavior-Induced Mirror-Prox TD Learning — Gradient TD methods for off-policy prediction with linear function approximation, tackling the geometry problem induced by auxiliary-variable metrics. If you’ve been burned by TD instability in practice, the Mirror-Prox angle is worth understanding.

  • Behavior-Aware Auxiliary Corrections for Off-Policy TD — Builds on TDC and TDRC with behavior-aware auxiliary covariance corrections for off-policy TD stabilization. Two papers on off-policy TD in one day — the field is clearly still working through this.

  • The Cognitive Categorical Transformer — 306M-parameter architecture augmenting GPT-2 Small with category-theoretic inductive biases and cognitive science inspirations. The “cognitive” branding is doing some heavy lifting, but the actual category-theory-meets-transformer approach is novel enough to merit a read. 📄

  • URIEL — Selective sustainable logging and post-harvest silvicultural treatment in tropical forests using airborne robotics. Not your typical ML paper, but a concrete application of robotics to an actual environmental problem. Refreshing.

  • Review Arcade — Examines human alignment and gameability of LLM-generated scientific paper reviews. Spoiler: if you think the peer review pipeline isn’t already being gamed by both reviewers and authors with LLMs, this will confirm your suspicions.

  • Orthogonal Concept Erasure for Diffusion Models — Addresses limitations in existing concept erasure methods for diffusion models, where training-based approaches are effective but computationally expensive. The orthogonal framing suggests a cleaner separation of concepts — if it works at scale, it’s meaningful.

AI Dev Tools

  • claude-mem — Captures everything an agent does during sessions, compresses it with AI, and injects relevant context into future sessions. Supports Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode, and more with Chromadb storage. Cross-agent persistent context is the dream — this is a step toward it.

  • Claude Code v2.1.154 — Opus 4.8 as default with high-effort reasoning, plus dynamic workflows that orchestrate tens to hundreds of background agents for complex tasks. Fast mode on Opus 4.8 too. The multi-agent orchestration angle is where things get genuinely architectural. 🔥

  • Goose v1.35.0 — Hooks system for pre/post tool execution callbacks, including PreToolUse denial hooks. Fine-grained control over agent tool usage — useful for safety checks, custom instrumentation, and generally keeping agents from doing dumb things. 🛠️

Today’s Synthesis

Two themes jump out today: context persistence and off-policy stability. Start with claude-mem and Claude Code v2.1.154 — together they sketch the shape of what serious agent workflows are becoming. claude-mem captures and compresses session context across agents (Claude Code, OpenCode, Gemini, etc.) using Chromadb, while Claude Code v2.1.154 pushes multi-agent orchestration with tens to hundreds of background agents and dynamic workflows. If you’re building anything beyond single-session tool calls, the combination of persistent compressed memory and parallel agent orchestration is the architecture to prototype now, not next quarter. On the theory side, the pair of off-policy TD papers — Behavior-Induced Mirror-Prox TD Learning and Behavior-Aware Auxiliary Corrections for Off-Policy TD — tackle the same stability problem from complementary angles: geometry-aware updates and covariance corrections. If your replay buffers are giving you divergent value estimates, both are worth implementing as baselines. The practical thread: context persistence and stable off-policy learning are converging on the same engineering question — how do you learn reliably from data you didn’t generate in the current session?