Sealevel — Solana Language Model

The Solana coding AI that actually knows Solana.

Sealevel(Solana Language Model) is a fine-tuned coding LLM specialized for Solana and Anchor development. Built on Qwen2.5-Coder-7B-Instruct (7B dense) with QLoRA on 270k Solana records, it scores 85% on a 80-task Solana/Anchor benchmark and 90% on HumanEval (general coding).

Use it in your browser, terminal, VS Code, or via Claude Code / Cursor / Windsurf through MCP.

---

Quick Start

Web

https://sealevel.tech

CLI

pip install sealevel
slm config --api-key sk-slm-xxx
slm chat "How do I derive a PDA in Anchor?"

VS Code

Install Sealevel - Solana Language Model from the Marketplace, or:

code --install-extension slm-vscode-0.1.0.vsix

Claude Code / Cursor / Windsurf (MCP)

claude mcp add --transport http slm-solana https://slm-mcp.run.app/mcp

API (OpenAI-compatible)

curl https://api.slm.dev/v1/chat/completions \
  -H "Authorization: Bearer sk-slm-xxx" \
  -H "Content-Type: application/json" \
  -d '{"model":"slm-solana","messages":[{"role":"user","content":"Write a token transfer in Anchor"}]}'

---

What It Does

• Chat — Ask Solana/Anchor questions, get accurate, modern-syntax answers
• Generate — Scaffold Anchor programs (escrow, staking, AMM, NFT, DAO)
• Review — Security scan for deprecated patterns, missing checks
• Migrate — Convert old Anchor code to 0.30+ patterns
• Explain Tx — Decode Solana transactions by signature
• Explain Error — Look up error codes (Token, Anchor, System, ATA)
• Autocomplete — Inline suggestions in VS Code

---

Architecture

     ┌─────────────────┐
     │   Developer     │
     └────┬────────────┘
          │
   ┌──────┴───────┬──────────┬──────────┐
   ▼              ▼          ▼          ▼
 Web UI       CLI         VS Code    MCP Server
 (Vercel)   (PyPI)      (Marketplace) (Cloud Run)
   │              │          │          │
   └──────┬───────┴──────────┴──────────┘
          ▼
    LiteLLM Proxy  (OpenAI-compatible, rate-limited)
          │
          ▼
        SGLang  ←── LoRA adapter (HF: WhyParabola/slm-solana-lora)
        (H100 on Akash, Qwen2.5-Coder-7B-Instruct base)
          ▲
          │
      RAG API  ←── Qdrant (latest Solana/Anchor docs)

---

Project Structure

Directory	Purpose
training/	CPT, SFT, DPO training scripts (Unsloth + PyTorch)
scripts/	38 data pipeline scripts (collection → dedup → filter → prepare)
data/	741K CPT records, 731K SFT records (dataset card)
synthetic/	14 template-based Q&A generators
deploy/	Akash SDL files, Dockerfiles for inference stack
slm-web/	Next.js 16 web app (chat, explainer, docs, dashboard)
slm-cli/	Python CLI (slm chat, slm gen, slm review, etc.)
slm-vscode/	VS Code extension (@slm chat, autocomplete)
slm-mcp/	MCP server (Claude Code, Cursor, Windsurf integration)
results/	Eval results + LoRA checkpoints

---

Benchmarks

Solana/Anchor (80 tasks)

Category	Score
PDA Derivation	14/15 (93%)
Anchor Constraints	15/15 (100%)
SPL Token Ops	10/10 (100%)
CPI Patterns	9/10 (90%)
Error Handling	9/10 (90%)
Transaction Construction	7/10 (70%)
Adversarial	4/10 (40%)
Overall	68/80 (85%)

HumanEval (general coding)

18/20 (90%) — fine-tuning preserved general programming ability.

Full results: results/phase1/eval_results.json

---

Deployment Status

Component	Status
Data pipeline	Live (741K records)
Model training	Phase 2 SFT complete (LoRA on HF)
Inference (SGLang + LiteLLM + Qdrant + RAG)	Live on Akash H100
Web app	Ready for Vercel (env vars needed)
CLI	Ready for PyPI
VS Code extension	Ready for Marketplace
MCP server	Ready for Cloud Run (HTTP transport)

---

Development

Run web app locally

cd slm-web
npm install --legacy-peer-deps
cp .env.example .env.local  # fill in your keys
npm run dev

Run CLI locally

cd slm-cli
pip install -e .
slm --help

Run MCP server locally

cd slm-mcp
npm install && npm run build
MCP_TRANSPORT=http PORT=8080 SLM_API_URL=https://... node dist/index.js

Train model (Akash H100)

See deploy/README.md for training container setup.

---

Contributing

• Issues / feature requests: GitHub Issues
• Pull requests welcome — run tests first (npm test / pytest)
• See component-specific READMEs for dev setup

---

License

MIT for all code.

Dataset note: Training data includes Lumo Labs datasets under AGPL-3.0. See DATASET_CARD.md for full license breakdown.

---

Acknowledgments

• Qwen Team — base model (Qwen2.5-Coder-7B-Instruct)
• Unsloth — 2x faster fine-tuning
• Solana Foundation — Sealevel RFP support
• Akash Network — decentralized GPU compute
• Lumo Labs — Solana-focused datasets