Tech.us small model Ethoric-10M tops Open SLM leaderboard in 10M category

Tech.us said its latest research model, Ethoric-10M, ranked first in the 10-million-parameter category on the Open SLM Leaderboard as of Oct. 27, 2026. The result is meant to sharpen the company’s AI consulting and implementation work as more customers look to smaller, task-specific models.

Why it matters: - Tech.us is using in-house model training to build practical engineering judgment for client AI projects. - The company says that experience is becoming more relevant as organizations move toward smaller, task-specific models for cost, latency and data-privacy reasons. - Ethoric-10M also provides a benchmark case for how architecture and staged training affect performance at very small scale.

What happened: - Tech.us announced that Ethoric-10M ranked No. 1 in the 10-million-parameter category on the Open SLM Leaderboard as of Oct. 27, 2026. - The small language model scored 41.60% on ArithMark-3, a multiple-choice elementary math benchmark focused on arithmetic reasoning. - The result was the highest score across the leaderboard’s 10M and 50M categories. - The same score would place fifth among 187 models in the 100M category.

The details: - Ethoric-10M is a 12-layer, decoder-only transformer with 9.97 million total parameters, including embedding parameters. - The model uses rotary position embeddings, grouped-query attention and SwiGLU activations. - Tech.us added four published architectural mechanisms to the base design. - Peri-LN applies RMSNorm before and after each attention and MLP sublayer to improve training stability. - Canon layers use short causal depthwise convolutions to mix each token with its three preceding tokens before the attention and MLP sublayers. - Per-head QK normalization rescales queries and keys within each attention head before attention scores are computed. - Headwise sigmoid gating controls how much of each head’s output passes through. - Ethoric-10M was trained on 19 billion tokens in two stages. - The first stage used 15 billion tokens of broad educational material. - The second stage used 4 billion tokens in a cooldown phase with a lower learning rate and a training mix shifted toward higher-quality mathematics, question-answer content and tutorial-style material. - The model was evaluated with zero-shot normalized accuracy under Open SLM Leaderboard methodology. - Ethoric-10M scored 59.85% on PIQA, 37.58% on ARC-Easy, 28.75% on HellaSwag and 24.66% on ARC-Challenge. - The HellaSwag and ARC-Challenge scores are near the 25% random-chance baseline for four-option tasks, which the company said is typical at this model size. - Ethoric-10M has a 512-token context window and was trained on English-language data. - The model has not been instruction-tuned or safety-aligned, and its outputs may be inaccurate or incoherent.

Between the lines: - Tech.us is treating Ethoric-10M as a research asset, not a product to deploy in customer systems. - The company says direct model-building gives its engineers hands-on practice with evaluation, training-data curation and deployment tradeoffs. - Tech.us cited an April 2025 Gartner prediction that organizations will use small, task-specific AI models at least three times more than general-purpose large language models by 2027. - Ethoric-10M extends the company’s earlier Tokle-3M research model, which was released on Hugging Face and examined guided training in a model with fewer than 3 million parameters. - Unlike Tokle-3M, Ethoric-10M explores a different mix of architectural mechanisms and staged training.

What's next: - Ethoric-10M is available on Hugging Face under the MIT license. - The research page includes the model’s architecture, training methodology, benchmark results and limitations. - Tech.us said it will continue using independent model research to inform its AI consulting, AI implementation and AI-augmented software development work.

The bottom line: - Tech.us wants Ethoric-10M to serve as both a leaderboard result and a training ground for the engineers who build customer AI systems.