---

Current Focus

• Nemostein — Nemotron/Nemotron-Omni reasoning and tool-use fine-tunes, including GGUF releases for local inference
• Ornstein — multimodal and MoE reasoning models (Qwen/Gemma-derived) optimized for local deployment and agentic workflows
• CRUCIBLE — Conditional Reward-Utility Coreset by Iterative Bayesian Learnability Estimation, a private data pipeline for reasoning and agent-trace curation
• SABER — spectral refusal-boundary editing with KLD drift measurement for capability-preserving behavior edits
• RYS — a separate post-training method line for refusal/behavior shaping experiments
• Acta / Talos — curated agentic tool-use and coding-assistant traces for SFT and evaluation
• Local inference — GGUF, quantized, and deployment-friendly builds for llama.cpp and MLX workflows

Featured Model Lines

Nemostein

Nemotron 3 Nano / Nemotron Omni reasoning line focused on chat reasoning, Hermes-style tool calls, and local deployment.

Nemostein-3-Nano-Omni-30b-a3b — Nemotron 3 Nano Omni lineage fine-tune for reasoning and tool-use chat

• BF16 / Transformers · GGUF
• GGUF variants: Q8_0, Q6_K, Q4_K_M, Q2_K
• Follow-on Acta Hermes tool-call SFT and SABER/KLD tuning are in progress.

Ornstein

Multimodal and MoE reasoning line across 27B–35B scales, shipped in formats for every common local-inference stack.

Ornstein-3.6-27B — multimodal base (Qwen 3.6 27B)

• Base · GGUF

Ornstein-Hermes-3.6-27b — Hermes-format function-calling fine-tune (multimodal, tool-use)

• Base · GGUF
• MLX: bf16 · 8-bit · 6-bit · 4-bit

Ornstein3.6-35B-A3B — text-only Mixture-of-Experts (34.7B total, ~3B active)

• Base · GGUF
• MLX: bf16 · 8-bit · 6-bit · 4-bit

Ornstein-27B-v2 — earlier 27B revision

• GGUF

CRUCIBLE

CRUCIBLE is a private data pipeline: Conditional Reward-Utility Coreset by Iterative Bayesian Learnability Estimation. It is used for reasoning-oriented data selection, self-correction supervision, and agent-trace curation.

SABER

SABER is a spectral refusal-boundary editing toolkit that tunes candidates against generated refusal rate and KLD drift to preserve useful capability.

• SABER code
• DJLougen/gemma-4-E2B-it-saber

RYS

RYS is a distinct post-training method line for refusal and behavior shaping experiments, tracked separately from SABER.

Code & Agent Infrastructure

• SABER — refusal-boundary and KLD-drift tuning toolkit for capability-preserving behavior edits
• RBForge — runtime tool creation for Ornstein, Hermes, and SABER agents, with forged tools persisted into RBMEM memory
• Rust-Brain — Rust rbmem CLI and structured memory format for timestamp-protected sections, hierarchy, graph relations, Hermes integration, and Markdown sync
• MPKnet — bio-inspired CNN modeling the M/P/K pathways of the lateral geniculate nucleus; competitive accuracy with 52× fewer parameters than ResNet18, designed for edge/embedded deployment

Featured Datasets

• Ornstein Curated 100K — curriculum-sorted multi-domain reasoning data
• Hermes Agent Traces Filtered — quality-filtered agent reasoning traces
• Harmonic Reasoning v1 — compact synthetic reasoning data for math, code, and self-correction
• WittgenSite — prompt consistency benchmark for AI coding agents
• Acta — curated agentic tool-use conversations, including Hermes-style tool-call traces
• Talos Scenarios — agentic task scenarios for synthetic trace generation

Principles

• Ship open artifacts that people can inspect, run, and adapt
• Pair model releases with practical local inference formats (GGUF, MLX, etc.)
• Treat datasets as first-class research objects, not just training fuel
• Explore behavior editing while measuring capability drift and documenting limitations
• Build for researchers, tinkerers, and builders who want systems they can actually run locally

Contact

Open a discussion on any repository or reach out via DJLougen.