Talk to an AI Expert

TECH.US AI RESEARCH

Ethoric-10M: The #1-Ranked Small Language Model in the 10M Category

Ethoric-10M is a 9.97-million-parameter research small language model developed by Tech.us, an AI consulting, AI implementation and AI-augmented software development company. As of 27th Oct 2026, it ranks first in the 10-million-parameter category on the Open SLM Leaderboard, with an Intelligence Index of 12.94.

By the Tech.us AI Research Team Published 27th Oct 2026

#1 IN THE 10M CATEGORY
12.94 INTELLIGENCE INDEX
41.60% ARITHMARK-3

MODEL AT A GLANCE

Ethoric-10M Specifications and Key Facts

A quick reference to Ethoric-10M's architecture, training configuration, benchmark performance, and research availability.

Developer
Tech.us, San Jose, California
Model type
Decoder-only transformer small language model
Parameters
9.97 million total, including embedding parameters
Training Data
19 billion English-language tokens in two stages
Ranking
#1 in the 10M category, Open SLM Leaderboard, as of 27th Oct 2026
Intelligence Index
12.94
Instruction-tuned or safety-aligned
No
License
MIT
Released
27th Oct 2026
Download
Hugging Face

BENCHMARK COMPARISON

How Does Ethoric-10M Perform on Standard Benchmarks?

A side-by-side comparison of Ethoric-10M and the next three models in the 10M category, evaluated across reasoning, commonsense, and arithmetic benchmarks using zero-shot normalized accuracy and the Open SLM Leaderboard methodology.

Explore Ethoric-10M
Benchmark Ethoric-10M ForgePlex-M2.2-10M Heron1i-flash-lite-v2 Min-spark-1.6
HellaSwag 28.75% 29.19% 28.86% 28.52%
ARC-Easy 37.58% 39.77% 38.89% 38.30%
Arc-Challenge 24.66% 22.70% 23.38% 22.44%
PIQA 59.85% 59.79% 58.54% 59.14%
ArithMark-3 41.60% 37.40% 36.70% 36.90%
Intelligence Index 12.94 12.12 11.11 11.08
Ethoric-10M leads Top result from another model

Source: Open SLM Leaderboard, as of 27th Oct 2026. Bold marks the best score in each row. Random-chance baselines are 25% for HellaSwag, ARC-Easy, and ARC-Challenge and 50% for PIQA. Gaps under about 2 percentage points on PIQA, ARC-Easy, and ARC-Challenge are within the margin of error for test sets of this size.

MODEL OVERVIEW

How Is Ethoric-10M Built?

Ethoric-10M brings four architectural mechanisms together to explore training stability, local token processing, and attention behavior within a compact transformer. 

Ethoric-10M is a 12-layer, decoder-only transformer with 9.97 million total parameters, including embedding parameters. It uses Rotary Position Embeddings (RoPE), Grouped-Query Attention (GQA), and SwiGLU as its foundation. The four mechanisms below come from recent published research, credited in the sources at the end of this page.

Peri-LN

Applies RMSNorm before and after each attention and MLP sublayer to support training stability.

Canon Layers

Short causal depthwise convolutions that mix each token with its three preceding tokens before the attention and MLP sublayers.

Per-Head QK-Norm

Normalizes queries and keys within each attention head before attention scores are computed.

Headwise Sigmoid Gating

A sigmoid gate on each attention head controls how much of that head's output passes through.

TRAINING APPROACH

19 Billion Tokens Across Two Training Stages

Ethoric-10M moves from broad educational training data toward a more focused mixture of mathematics, question-answer, and tutorial-style content.

Stable Stage

A broad mix of educational web text, textbooks, question-answer data, and mathematics.

15B Tokens

Cooldown Stage

The learning rate decreases as the data mixture shifts toward higher-quality mathematics, QA, and tutorial-style material.

4B Tokens

ARITHMETIC PERFORMANCE

A 10M-Parameter Model. Competing With Models 5X Its Size.

Ethoric-10M achieved 41.60% on ArithMark-3, outperforming models in the 50M-parameter category despite having approximately one-fifth as many parameters. Its score ranks first across the 10M and 50M categories and would place fifth among 187 models in the 100M category.

The Role of Cooldown Training

Ethoric-10M's cooldown stage increased the share of mathematical data. Its arithmetic results highlight the potential of carefully structured training at a compact model size.

ArithMark-3 Accuracy

41.60%

#1

Highest score across the 10M & 50M categories

#5

Would place in the 100M category

RESEARCH FOCUS

What Ethoric-10M Is Designed to Explore

The model brings together architectural and training choices to explore how they influence learning and performance at a compact scale.

Attention Stability and Control

Examining how normalization and gating mechanisms shape the attention process.

Local Context Processing

Studying how short-range token mixing works alongside causal attention.

Structured Training

Exploring how changes in training data composition across stages influence the learning process.

We didn’t build Ethoric-10M to sell it. We built it to understand it. A 10-million-parameter model leaves nowhere to hide a weak architecture or a poor data mix, so every choice shows up in the scores. That’s the discipline we want our engineers bringing to client work.

Praveen Narra Founder & CEO, Tech.us

FOR TECH.US CLIENTS

What This Research Means for Tech.us Clients

Ethoric-10M will not be deployed in client systems. Its value to clients is the first-hand experience it gives the engineers who deliver Tech.us' AI consulting, AI implementation, and AI-augmented software development work.

3× By 2027

That experience matters as companies shift toward smaller models.

According to an April 2025 Gartner prediction, organizations will use small, task-specific AI models at least three times more than general-purpose large language models by 2027.

Training models from scratch gives Tech.us engineers direct practice with the decisions that determine whether those deployments succeed:

  • when a small model can replace a large hosted one on cost, latency, or data-privacy grounds
  • How to curate and stage training data
  • How to design evaluations that hold up to scrutiny.
In 26 years of building software, I’ve learned you can’t advise clients on technology you’ve never built yourself. AI is no different. Our team trained Ethoric-10M from scratch, so when we tell a client which model to use and how to deploy it, that advice comes from first-hand experience.
1-praveen-narra Praveen Narra Founder & CEO, Tech.us
Talk to Tech.us About AI Implementation

FAQ

Frequently Asked Questions About Ethoric-10M

Ethoric-10M is a research small language model with 9.97 million parameters, developed by Tech.us, an AI consulting, AI implementation and AI-augmented software development company in San Jose, California. It combines four published architectural mechanisms with a two-stage training process spanning 19 billion tokens. 

As of 27th Oct 2026, Ethoric-10M ranks first in the Open SLM Leaderboard's 10-million-parameter category, with an Intelligence Index of 12.94. Its 41.60% ArithMark-3 score is the highest of any model in the 10M and 50M categories and would place fifth among the 187 models in the 100M category. 

Ethoric-10M is a research model intended for experimentation and evaluation rather than deployment in live applications. It has a 512-token context window and was trained on English-language data, without instruction tuning or safety alignment. For production-ready AI solutions, Tech.us provides AI consulting and implementation services to help businesses move from pilot to production. 

Tech.us builds research models such as Ethoric-10M and Tokle-3M so its engineers gain first-hand experience with model design, training data, and evaluation. Tech.us applies that experience in its AI consulting, AI implementation, and AI-augmented software development work for mid-market companies in healthcare, construction, insurance, transportation, and manufacturing. 

A small, task-specific model is often the better fit when the task is narrow, responses must be fast, inference cost matters, or data must stay on infrastructure the company controls. Large general-purpose models remain stronger for open-ended tasks. Gartner predicts that by 2027, organizations will use small, task-specific AI models at least three times more than general-purpose LLMs. 

Ethoric-10M is available on Hugging Face under the MIT license. The Hugging Face model card lists the architecture, training stages, benchmark results, and known limitations.

Tokle-3M, Tech.us' earlier model with fewer than 3 million parameters, tested a guided training method that adds information about token relationships during training. Ethoric-10M, with 9.97 million parameters, tests a different approach: four architectural mechanisms combined with a staged training schedule. Tokle-3M is available on Hugging Face.

Explore the Research Behind Ethoric-10M

Ethoric-10M brings Peri-LN normalization, local token mixing, attention gating, and staged training together within a 9.97M research model.

Explore its architecture, benchmark results, and technical configuration.

Download Ethoric-10M on Hugging Face

Sources

  1. Open SLM Leaderboard, as of 27th Oct 2026
  2. Kim et al. (2025), Peri-LN normalization
  3. Allen-Zhu (2025), Physics of Language Models: Part 4.1, Architecture Design and the Magic of Canon Layers
  4. Henry et al. (2020), Query-Key Normalization for Transformer
  5. Qiu et al. (2025), Gated Attention for Large Language Models
  6. Gartner Predicts by 2027, Organizations Will Use Small, Task-Specific AI Models Three Times More Than General-Purpose Large Language Models, April 9, 2025