AmeforgeAmeforge

Foundation Model · Reasoning Variant · 2026

SAM-G-Reasoning

A reasoning-augmented variant of SAM-G. Structured multi-step inference at 30.3M parameters — logic, planning, and procedural chains, fully on-device.

77.8%
Aggregate exact match
100%
Logic & plan chain
30.3M
Parameters (fp32: 121MB)
Offline
No cloud dependency

Overview

Structured reasoning at the edge

SAM-G-Reasoning extends the SAM-G architecture with explicit multi-step reasoning capabilities. Where the base model excels at action generation and structured output, the reasoning variant adds procedural chain inference — enabling it to handle logic, planning sequences, and unit conversions without external tools or cloud APIs.

The model achieves 100% exact match on logic (modus ponens, modus tollens, chained syllogisms) and multi-step planning tasks. These results hold at 30.3M parameters — a fraction of the size of comparable open models that achieve lower scores.

Like all Ameforge models, SAM-G-Reasoning is designed for offline, on-device deployment. It runs without internet connectivity, without cloud inference, and without the energy overhead of large-scale models.

Benchmark

Reasoning evaluation by domain

Task familyTypeExact match
Logic (ponens/tollens/chains)Procedural100%
Plan chain (multi-step actions)Procedural100%
Conversion (unit chains)Procedural100%
Sequence (next term)Procedural100%
Date/time (clock/calendar)Procedural92%
Compare (max/min)Procedural92%
State tracking (device toggles)Working memory83%
Parity/digitsMixed58%
Count/filterCalculation67%
Sort listCalculation50%
Word problemCalculation50%
Arithmetic chainCalculation42%
Aggregate77.8%

Comparison

Efficiency vs. comparable models

ModelParametersAggregate exact match
SAM-G-ReasoningOurs30.3M77.8%
Pythia-70M70M~12%
SmolLM2-360M362M~28%
Qwen2.5-0.5B494M~34%
Qwen2.5-1.5B889M~41%

Aggregate exact match across all 12 reasoning task families. Baseline figures are approximate.

Capabilities

What SAM-G-Reasoning does

Logical inference

Modus ponens, modus tollens, and chained syllogisms. The model handles formal logic structures with 100% exact match, making it suitable for rule-based decision systems.

Multi-step planning

Plan chain tasks requiring sequences of dependent actions. The model resolves action dependencies and outputs ordered, structured plans without external orchestration.

Procedural chains

Unit conversions, date/time calculations, and sequence prediction. These procedural tasks require step-by-step computation that the model handles natively.

State tracking

Device toggle sequences and working memory tasks. The model maintains intermediate state across reasoning steps, enabling it to track changes over multi-step inputs.

Available now

Download SAM-G-Reasoning

Available on HuggingFace. fp32 checkpoint, 121MB. No API key required.

HuggingFace — AMFORGE/SAM-G-Reasoning