TLM · Manufacturing v1 · 1.7B

FR-Forge-1.7B

A Thin Language Model for the factory floor. Grounded reasoning across operations, quality, and supply chain — running locally on Apple Silicon.

Parameters
1.7B
Quantization
4-bit
Context
32K
Runtime
MLX
License
Apache-2.0
Capabilities

Built for the factory floor

FR-Forge covers the reasoning tasks that matter in manufacturing — without sending your operational data to a cloud.

001

Operations Reasoning

Interpret production metrics, flag throughput anomalies, and surface root causes from machine logs and shift reports in natural language.

002

Quality Analysis

Parse inspection records, correlate defect patterns with process parameters, and draft corrective action summaries grounded in your data.

003

Supply Chain Intelligence

Reason over supplier data, lead times, and inventory levels to surface risks and recommend reorder decisions with traceable logic.

004

Maintenance Forecasting

Analyse equipment telemetry and maintenance history to predict failure windows and prioritise preventive interventions.

005

Document Understanding

Extract structured data from SOPs, work orders, and compliance documents — no cloud, no latency, no data leaving the site.

006

Multi-step Planning

Decompose complex production scheduling problems into ordered steps, with self-correction when constraints shift mid-plan.

Architecture

How FR-Forge works

Distillation from frontier reasoning into a weight set that fits on a MacBook Pro. No cloud, no API key, no latency.

01

Domain Corpus

Trained on a proprietary corpus of manufacturing SOPs, maintenance logs, quality records, supply chain data, and industrial domain literature — curated and reviewed in-house.

02

Frontier Distillation

We use frontier models as teachers, extracting reasoning chains specific to manufacturing contexts. The student model learns to replicate the judgment, not just the vocabulary.

03

LoRA Fine-tuning via MLX

Efficient parameter fine-tuning on Apple Silicon with MLX, producing a 4-bit quantized model that runs at inference speeds practical for real-time operational use.

04

On-device Inference

Runs locally. Your operational data stays on your machine. No egress, no rate limits, no subscription — integrate via the MLX Python API or our REST wrapper.

Benchmarks

Performance against larger baselines

Evaluated on manufacturing-domain reasoning tasks. FR-Forge outperforms models 10× its size on domain-specific benchmarks.

Task FR-Forge 1.7B GPT-4o-mini Llama-3.1 8B
Quality defect classification 91.4% 84.2% 78.6%
Root cause extraction (F1) 88.7 81.3 74.1
SOP comprehension accuracy 94.1% 89.5% 83.2%
Supply chain risk scoring 87.9% 79.8% 71.4%
Maintenance window prediction 83.2% 76.1% 68.9%
Hardware Requirements

Runs on what you already have

Apple Silicon

Recommended · MLX runtime
ChipM1 or later
RAM8 GB unified (16 GB recommended)
Storage~1.1 GB (4-bit)
Tokens/sec~42 tok/s on M2 Pro

CUDA GPU

Supported · llama.cpp / GGUF
GPU VRAM4 GB minimum
RAM8 GB system RAM
Storage~1.1 GB (Q4_K_M)
Tokens/sec~60 tok/s on RTX 3080

Deploy FR-Forge
on your factory floor.

Download the weights, integrate via the API, or commission a bespoke version trained on your proprietary data.