TECH.US AI RESEARCH
Tokle-3M ranks first among models under 3 million parameters on the Open SLM Leaderboard , as of 1st Oct 2026.
Developed by Tech.us as a research model, Tokle-3M has 2.91 million parameters and uses SPAB, a guided training approach that introduces additional information about relationships between tokens during training. The work explores how training methods, not just model size, influence how efficiently a model learns and performs.
BENCHMARK COMPARISON
A side-by-side view of Tokle-3M results across reasoning, commonsense, arithmetic, and the composite Intelligence Index.
Try Tokle-3M →| Benchmark | Tokle-3M 2.91M | Pulvis-v2 2.96Mx3* | Purrence-3M 2.99Mx6* | Ember-2 2.96Mx2* |
|---|---|---|---|---|
| HellaSwag | 27.20% | 27.93% | 27.81% | 27.28% |
| Arc-Easy | 34.85% | 31.86% | 34.55% | 33.42% |
| Arc-Challenge | 23.98% | 22.78% | 23.63% | 22.01% |
| Piqa | 55.01% | 56.86% | 56.64% | 55.11% |
| ArithMark-3 | 40.80% | 37.40% | 34.80% | 35.90% |
| Intelligence Index | 8.91 | 8.62 | 8.49 | 7.21 |
RESEARCH OBJECTIVE
Tokle-3M was developed to study how architecture and training methods influence model performance when compactness is an intentional design choice.
The research focuses on how a purpose-built model can make effective use of what it learns. Rather than relying on model size alone, Tech.us is studying how additional guidance during training can help a model identify useful relationships and retain that learning.
We designed this model to be purpose-built from the start, with guided training helping it make better use of what it learns. The focus is not on making it broader, but on making it more deliberate, efficient and useful for defined tasks.
Praveen Narra
Founder & CEO, Tech.us
TRAINING APPROACH
Tokle-3M grew out of Tech.us’ work on Static Pairwise Attention Bias (SPAB), a training approach designed to give the model additional information about relationships between tokens while it learns.
The process happens in two stages.
SPAB uses statistical patterns in the training data to identify words or tokens that are more likely to be related.
That information is added to the model's attention process during training, helping it focus on useful relationships earlier rather than having to discover all of them on its own.
The additional SPAB information is then removed. What the model learned with that guidance is distilled into the final Tokle-3M model.
The final model runs with approximately 2.91 million parameters without requiring the additional SPAB table at inference.
BENCHMARKS
Tokle-3M was evaluated across five established benchmarks using zero-shot normalized accuracy and the Open SLM Leaderboard methodology.
The Open SLM Leaderboard's Intelligence Index combines chance-adjusted results across reasoning, commonsense and arithmetic benchmarks into a composite measure.
MODEL PROFILE
A concise view of the model’s architecture, training approach, evaluation method, and current research scope.
Tokle-3M is not instruction-tuned or safety-aligned and is currently positioned as a research model rather than a production assistant.
Because the model uses Rotary Position Embeddings (RoPE), its context window can be extended at inference time without changing the architecture.
| Specification | Detail |
|---|---|
| Parameters | Approximately 2.91 million |
| Training Data | Pre-trained on English datasets |
| Position Embeddings | Rotary Position Embeddings (RoPE) |
| Training Approach | SPAB-assisted guided training |
| Evaluation Methodology | Open SLM Leaderboard methodology |
| Benchmark Method | Zero-shot normalized accuracy |
| Current Scope | Research model |
RESEARCH FINDINGS
Tokle-3M gives Tech.us a focused way to examine how specific training and architectural decisions affect model performance.
The SPAB-assisted training process allows Tech.us to study how providing additional information about token relationships influences what the model learns and how much of that learning remains after the additional guidance is removed.
Tokle-3M was designed as a compact model from the start. This allows the research to focus on how architecture and training choices contribute to performance within a deliberately defined model design.
The research also supports Tech.us’ exploration of models designed around defined tasks, domains and operating requirements rather than maximum breadth.
DEPLOYMENT
Tokle-3M supports Tech.us’ exploration of deployment approaches that give organizations greater control over where models run and how information is handled.
Explore deployment closer to the point of use.
Explore deployment within infrastructure the organization manages.
Support approaches that keep more processing within controlled environments.
Tokle-3M is one example of how Tech.us is exploring more focused approaches to AI. If your organization is considering Small Language Models for a specific workflow, deployment environment or data requirement, talk with our team about where an SLM approach could make sense.
Talk to Our AI Expert →We value your privacy
By continuing to use this website, you agree to our Privacy Policy.
If you decline, your information won’t be tracked when you visit this website. A single cookie will be used in your browser to remember your preference not to be tracked.
Necessary cookies keep the site running and are always on. Turn the others on or off to control how Tech.us uses them.
Required for the site to function. Cannot be disabled.
Help us measure traffic and see how visitors use the site so we can improve it. All information is aggregated.
Used to deliver and personalize ads, measure campaign performance, and share data with advertising partners (including Retention.com and RB2B).
For residents of California and other states with similar rights. Turning this on opts you out of the sale or sharing of your personal information for targeted advertising (this also disables Advertising cookies).
Learn more in our Cookie Policy and Privacy Policy. You can change these settings at any time.
