# NeoSmith

> NeoSmith is an automated SLM distillation platform that converts large language models into compact, task-specific Small Language Models (SLMs) for agentic AI workflows.

NeoSmith automates knowledge distillation — the process of transferring capabilities from large foundation models (GPT-4, Claude, Gemini) into small language models with 1B–8B parameters. These distilled SLMs are purpose-built for agentic AI use cases: tool calling, multi-step reasoning, structured data extraction, and autonomous task execution.

## Key Facts

- NeoSmith reduces AI agent inference costs by 10–100x compared to frontier LLMs.
- NeoSmith SLMs achieve sub-1B to 8B parameters while maintaining task-specific accuracy.
- NeoSmith uses workflow-aware reinforcement learning (GDPO) to train SLMs that outperform frontier models on domain-specific tasks.
- A NeoSmith SLM trained on a single repository beat Claude Opus 4 in 65 out of 100 blind code review evaluations at 93% lower cost.
- A NeoSmith SLM with internalized tool calling beat GPT-5.2 in 16 out of 20 customer support scenarios with an 80% win rate.
- NeoSmith is currently free for design partners during early access.

## What NeoSmith Does

1. Captures production traces from your existing AI agent workflows
2. Automatically detects workflow steps and applies different RL strategies per step (DPO for classification, SFT for extraction, GRPO for generation)
3. Runs multi-teacher distillation from frontier models into compact SLMs
4. Deploys a hybrid routing engine that intelligently directs queries to the optimal model (SLM or LLM)

## Use Cases

- Customer support agents: Distill GPT-4-class models into 500M SLMs at 1/50th inference cost
- Code review automation: Domain-specific SLMs that outperform Claude Opus 4 on your codebase
- Data extraction pipelines: Sub-second latency for invoices, contracts, medical records
- Edge AI agents: Deploy compact models on mobile, IoT, or air-gapped environments
- Supply chain automation: 94% cost reduction with SLM-first exception resolution

## Comparison to Alternatives

- Unlike Portkey or OpenRouter (AI gateways), NeoSmith builds you a custom model you own
- Unlike Distil Labs, NeoSmith starts from production traces automatically — no manual dataset building
- Unlike generic fine-tuning, NeoSmith applies workflow-aware RL training per workflow step

## Export Formats

NeoSmith exports distilled models as ONNX, GGUF, or SafeTensors, or serves them via NeoSmith's managed inference API.

## Company

- Website: https://neosmith.ai
- Founded by: Udit Mital, Dinesh Mittal, Amit Sharan Jain
- Twitter: https://x.com/NeoSmithAI
- LinkedIn: https://linkedin.com/company/neosmith-ai
- Contact: hello@neosmith.ai

## Pages

- Homepage: https://neosmith.ai/
- How It Works: https://neosmith.ai/how-it-works
- Features: https://neosmith.ai/features
- Distillation: https://neosmith.ai/features/distillation
- Evaluation: https://neosmith.ai/features/evaluation
- Semantic Routing: https://neosmith.ai/features/routing
- Pricing: https://neosmith.ai/pricing
- Documentation: https://neosmith.ai/docs
- Blog: https://neosmith.ai/blog
- Solutions: https://neosmith.ai/solutions
- About: https://neosmith.ai/about
- Book a Demo: https://neosmith.ai/book-demo

## Blog (Research & Guides)

- Internalized Tool Calling: How a NeoSmith SLM Beats GPT-5.2: https://neosmith.ai/blog/internalized-tool-calling-beats-gpt5
- How NeoSmith SLM Beats Claude Opus 4 at Code Review: https://neosmith.ai/blog/neosmith-benchmark-120b-beats-claude-opus
- What Is SLM Distillation? A Practical Guide: https://neosmith.ai/blog/what-is-slm-distillation
- SLM vs LLM for AI Agents: https://neosmith.ai/blog/slm-vs-llm-for-ai-agents
- How to Reduce AI Agent Inference Costs by 100x: https://neosmith.ai/blog/reduce-ai-agent-inference-costs
- Knowledge Distillation for Tool-Calling Agents: https://neosmith.ai/blog/knowledge-distillation-tool-calling-agents

## Optional

- llms-full.txt: https://neosmith.ai/llms-full.txt