June 17, 2026

Agentic Frameworks, Tooling, Skills

• ProvenanceGuard, https://huggingface.co/papers/provenanceguard-source-aware-factuality-verification-for-mcp-based-llm-agents — Runtime provenance tracking and hallucination filtering specifically for MCP tool calls; stabilizes external retrieval loops and prevents semantic drift in your research pipeline without requiring full model retraining. [Source: https://huggingface.co/papers/provenanceguard-source-aware-factuality-verification-for-mcp-based-llm-agents]
• PreAct, https://arxiv.org/abs/2606.17929 — Introduces dynamic shortcut routing for iterative agentic subroutines; cuts token waste and inference latency on repetitive rover telemetry parsing or vision-scraper scraping loops by collapsing redundant reasoning traces. [Source: https://arxiv.org/abs/2606.17929]
• FlowRAG, https://arxiv.org/abs/2606.17856 — Replaces rigid vector retrieval with frequency-aware multi-granularity graph flow routing; directly upgrades your cross-document dependency tracing in research pipelines without the latency overhead of dense RAG pipelines. [Source: https://arxiv.org/abs/2606.17856]
• OPD-Evolver, https://huggingface.co/papers/opd-evolver-cultivating-holistic-agent-evolver-via-on-policy-distillation — Automates iterative refinement of agent skill weights via on-policy feedback; reduces manual prompt engineering overhead when stitching multi-stage workflow chains. [Source: https://huggingface.co/papers/opd-evolver-cultivating-holistic-agent-evolver-via-on-policy-distillation]

Notable Research

• ACE-Ego-0, https://huggingface.co/papers/ace-ego-0-unifying-egocentric-human-and-robotic-data-for-vla-pretraining — CUHK/VLA pretraining dataset bridging human egocentric views with robotic actuation; provides direct grounding patterns for rover locomotion and aerospace inspection telemetry without heavy sim-to-real fine-tuning. [Source: https://huggingface.co/papers/ace-ego-0-unifying-egocentric-human-and-robotic-data-for-vla-pretraining]
• ActWorld, https://huggingface.co/papers/actworld-from-explorable-to-interactive-world-model-via-action-aware-memory — ByteDance's action-aware world model architecture; enables predictive environment simulation for rover navigation loops without full physical training runs. [Source: https://huggingface.co/papers/actworld-from-explorable-to-interactive-world-model-via-action-aware-memory]
• EvolveNav, https://arxiv.org/abs/2606.18235 — Adds self-evolving memory to zero-shot object navigation; directly applicable to stabilizing Kangaroo rover spatial reasoning in dynamic/unmapped terrain where static waypoints fail. [Source: https://arxiv.org/abs/2606.18235]
• Fixed-Point Reasoners, https://huggingface.co/papers/fixed-point-reasoners-stable-and-adaptive-deep-looped-transformers — Stabilizes recursive reasoning loops by converging attention weights to fixed points; reduces divergence in long-horizon agentic planning without gradient updates, useful for extended rover decision cycles. [Source: https://huggingface.co/papers/fixed-point-reasoners-stable-and-adaptive-deep-looped-transformers]

Frontier Lab Updates

Nothing new today.

Models to Download & Try

• microsoft/FastContext-1.0-4B-SFT, https://huggingface.co/microsoft/FastContext-1.0-4B-SFT — 4B model. ~2–3GB VRAM (Q4_K_M). Leaves >28GB headroom for 131k context or heavy agentic buffers. New model release for yesterday's routing paper; trained context routing replaces heavy vector RAG in coding agents, ideal for lightweight research pipeline routing. [Source: https://huggingface.co/microsoft/FastContext-1.0-4B-SFT]
• unsloth/diffusiongemma-26B-A4B-it-GGUF, https://huggingface.co/unsloth/diffusiongemma-26B-A4B-it-GGUF — 25B model. ~13–14GB VRAM (Q4_K_M). Updated GGUF variant of the 26B vision tokenizer; directly advances local vision-scraper grounding and cross-modal feature routing for aerospace telemetry without re-downloading base weights. [Source: https://huggingface.co/unsloth/diffusiongemma-26B-A4B-it-GGUF]
• yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF, https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF — 12B model. ~6–7GB VRAM (Q4_K_M). High-efficiency coding/fable-composer variant; fits alongside other models for multi-agent tool routing or fallback reasoning on your 32GB setup. [Source: https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF]
• lordx64/Qwable-v1, https://huggingface.co/lordx64/Qwable-v1 — 36B model. ~19–20GB VRAM (Q4_K_M). Direct Qwen 3.6:35b successor/competitor; claims improved architectural efficiency and reasoning stability. KV cache for 131k tokens at fp8 consumes ~9.5GB, totaling ~29.5GB on 32GB VRAM. Positions itself as a lighter 32GB-friendly alternative with upgraded context handling for agentic synthesis. [Source: https://huggingface.co/lordx64/Qwable-v1]

Skipped as Already Covered

• MiniMaxAI/MiniMax-M3 (427B parameter count, exceeds 32GB hardware limit)
• moonshotai/Kimi-K2.7-Code (1.1T parameter count, exceeds 32GB hardware limit)
• zai-org/GLM-5.2 (753B parameter count, exceeds 32GB hardware limit)
• deepseek-ai/DeepSeek-V4-Pro (862B parameter count, exceeds 32GB hardware limit)
• Anthropic Fable 5/Mythos 5 export suspension (visible safeguards/export control, covered 6/12/6/13)
• Qwopus3.6-27B-Coder-MTP-GGUF (agentic inference routing, covered 6/15)