Tenkai Daily — July 16, 2026

Model Releases

• thinkingmachines/Inkling — A multimodal MoE model handling image/audio/text-to-text conversations, Apache-2.0 and US-region. Another entry in the “make the model eat everything” sweepstakes 🤖
• ATH-MaaS/OvisOCR2 — Qwen3.5-0.8B fine-tune built for OCR and document parsing, including tables and formulas to markdown, with vLLM support. Actually useful if your pipeline still chokes on scanned PDFs 📄
• AngelSlim/Hy3-GGUF — Quantized GGUF build of Tencent’s Hy3 with imatrix, ready for local text generation. For the folks running models on whatever spare GPU they can find 🛠️

Research Worth Reading

• OriginBlame: Record- and Token-Level Data Provenance for AI Training Datasets — Tackles the unlearning headache: locating exactly which training records belong to a given author when they ask for removal. Practical plumbing work most papers pretend doesn’t exist.
• SPINE: Bridging the Cyber-Physical Gap with Agentic AI — Argues the real robot deployment bottleneck isn’t the brain, it’s the “spinal cord” — calibration and platform integration. Worth a look if you’ve ever watched a demo fall over in real life.
• Interventional Grounding Audits: Black-Box Premise-Dependency Tests for LLM Chain-of-Thought via Predicate Substitution — Black-box method to check whether CoT actually depends on its stated premises or just hallucinates a logical trail. The “is it thinking or performatively thinking” test 🔥
• Probabilistic Extension of Neuro-Symbolic AGI Robots based on Belnap’s Typed Intensional FOL — Extends neuro-symbolic IFOL_B with probability for self-referential reasoning in robots. The title alone should filter your reading queue.
• Self-Improvements in Modern Agentic Systems: A Survey — Frames the shift from research prototypes to deployed self-improving agents around controllable evolution with minimal human input. Good map if you’re being asked to “just make the agents learn on their own” at work.
• Improving Molecular Property Prediction in Small Language Models Using Graph-based Tools — Adds graph-based tooling to small LMs to fix their structural blindness on SMILES strings. Niche, but a real fix for a real SLM limitation.

Today’s Synthesis

The throughline today is that the gap between a model and a working system is where the real engineering lives. SPINE makes the point bluntly: robot demos don’t fail because the model is dumb, they fail because calibration and platform glue — the “spinal cord” — is missing. That’s the same lesson buried in OvisOCR2 : a fine-tune on Qwen3.5-0.8B only matters because it ships vLLM support and actually parses tables to markdown instead of hand-waving the integration. Meanwhile Interventional Grounding Audits gives you a black-box way to check if your CoT is reasoning or just narrating — directly useful before you trust an agent in that fragile integration layer. Concrete takeaway: stop evaluating models in isolation. If you’re dropping thinkingmachines/Inkling or a local Hy3-GGUF build into a pipeline, budget more time for the spinal cord than the brain, and run a premise-dependency audit on anything claiming to reason before it touches production.