AI & ML interests: Open reasoning models for local deployment. Post-training, agentic tool use, psychometrics-grounded evaluation, and efficient inference (GGUF, MLX, NVFP4).

Gestalt Labs 🇨🇦

Independent Canadian open reasoning research

Gestalt Labs builds and releases fully permissive open-weight reasoning systems: multimodal assistants, tool-using agents, and local-first deployments. Our methods are grounded in psychophysics and measurement science, applying signal detection theory and psychometric modeling to post-training and evaluation.

Models · Datasets · GitHub

Methods

• NSC-ACE — contrastive steering directions extracted from hidden states, refined with GRPO rollouts.
• SABER — capability-preserving refusal shaping for local agents.
• Curated data pipelines — DDM-inspired curation achieving AUC 0.97 with 53% token savings at 99.5% sensitivity.

Model lines

Ornstein 3.6 — multimodal and MoE reasoning models (27B–35B), shipped in GGUF, MLX, and full-precision formats for every common local stack.

• Ornstein-Hermes-3.6-27B-GGUF — multimodal, Hermes-format tool calling
• Ornstein-Hermes-3.6-27B-SABER-GGUF — Hermes + SABER
• Ornstein-3.6-27B-GGUF — multimodal base
• Ornstein3.6-35B-A3B — multimodal MoE, ~3B active (GGUF)

Ornstein 3.5 (9B) — compact multimodal models for local and edge stacks.

• Ornstein-3.5-9B-V2 — (GGUF, NVFP4)
• Ornstein-9Bv2.5-NSC-ACE — NSC-ACE + GRPO tool calling (GGUF)

Harmonic — compact Hermes-format reasoning models with structural output supervision.

• Harmonic-Hermes-9B-MLX

Software

• hive — unified agent memory and context compression
• honey-comb — CPU-only inline context compression
• Rust-Brain — structured agent memory (.rbmem)

Architectures

• MPKx — LGN-inspired vision with stride-based M/P/K pathways (874K params; 15.5K / 33 FPS on Raspberry Pi; competitive with ResNet-18 at 52× fewer parameters). The same repo also ships Kuramoto-VLM: coupled Kuramoto oscillators as the vision head of a frozen Qwen3.5-2B.

Featured datasets

• Hermes Agent Traces Filtered — 3,679 quality-filtered agent trajectories
• Ornstein Curated 100K — curriculum-sorted multi-domain reasoning data
• Acta — curated agentic tool-use conversations
• Phone AI Contention Bench — on-device agent evaluation seed

Principles

• Ship open artifacts people can inspect, run, and adapt without restriction
• Pair every release with practical local formats (GGUF, MLX, NVFP4)
• Treat datasets as first-class research objects

Contact

Open a discussion on any repository