viku-large

<p align="center"> <img src="https://huggingface.co/geethudinohtg/viku-large/resolve/main/logo.png" width="170" alt="Viku dino"/> <br/> <b>Viku Large</b> — Your Local Coding Captain &nbsp;·&nbsp; 4.4GB &nbsp;·&nbsp; 8GB Laptops &nbsp;·&nbsp; Apache 2.0<br/> <a href="https://huggingface.co/geethudinohtg/viku-large"><img alt="HF" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-viku--large-6095d6?style=flat"></a> <a href="#-quick-start"><img alt="LM Studio" src="https://img.shields.io/badge/LM%20Studio-Discover-181a1e?style=flat"></a> <a href="#-quick-start"><img alt="Ollama" src="https://img.shields.io/badge/Ollama-viku--large-000000?style=flat"></a> </p>

Not just Qwen with a new name. Viku is Qwen's engine with a completely different system around it — tools, memory, and a captain's judgment trained on 19.5K agentic traces. Same horsepower, different ship.

Why Viku, not plain Qwen?

---

✨ What you can build — 6 real use cases

# Viku's code style (what it actually writes):
def retry_with_backoff(fn, max_attempts=5, base_delay=0.5):
    """Exponential backoff + jitter — Viku never does silent except: pass"""
    import random, time
    for attempt in range(1, max_attempts + 1):
        try:
            return fn()
        except Exception as exc:
            if attempt == max_attempts: raise
            delay = min(30.0, base_delay * 2**(attempt-1)) * random.uniform(0.5, 1.5)
            print(f"attempt {attempt} failed ({exc}); retrying in {delay:.1f}s")
            time.sleep(delay)

📊 Honest Benchmarks

We A/B every release on the same 15-problem harness (evals/run_eval.py:1). If Large doesn't beat v1, we don't ship it as stable. No cherry-picking.

🧠 Under the Hood

flowchart LR
  Client[Any Agent<br/>opencode / Cline / aider] --> Proxy[memtree proxy :8787<br/>memory + router + tools]
  Proxy -->|code/scrape| Viku[Viku Large<br/>Q4 4.4GB]
  Proxy -->|deep reasoning| Teacher[DeepSeek-16B<br/>local]
  Viku --> Tools[web_scrape · http_fetch<br/>read/write files · memory]
  Tools --> Verbatim[VERBATIM EXTRACT<br/>never hallucinate DOM]

• Base: Qwen2.5-Coder-7B-Instruct (Apache 2.0) — thank you, Qwen team
• Method: QLoRA r=16, 4-bit, Unsloth, seq 2048, 19.5K ChatML samples
• Data: Magicoder (OSS-Instruct) + Evol-Instruct + Glaive/Hermes function-calling + 317 distilled traces (DeepSeek-16B, free local)
• Context: 32K native → YaRN 128K in v3 (viku-large is the "Large" size — like Starbucks Grande. Next: Venti = MoE)

🚀 Quick Start

LM Studio (friends: one search): Discover → search geethudinohtg/viku-large → Download → Load

Ollama:

ollama create viku-large -f Modelfile.large
ollama run viku-large

Any OpenAI client + full power (tools + memory + routing):

python -m memtree.serve   # http://localhost:8787/v1
# point client at http://localhost:8787/v1, model "viku-large"
# add "viku_native_tools": true → captain gets his hands

Large files or codebases → feed the memory directly:

curl -X POST http://localhost:8787/memory/ingest -H "Content-Type: application/json" \
  -d '{"path": "./huge_repo", "title": "my codebase"}'

⚠️ Limitations (honest)

• 7B-class: not a frontier 70B — great at coding/automation, weaker at deep math vs 30B+ models
• Identity can still wobble on creative phrasings (proxy lock + v3 armor fix this permanently)
• Tool-calling in v1 is via deterministic auto-scrape fallback — v2+ does native OpenAI tool-calls

🗺️ Roadmap

viku-large (Large, now) → viku-v3 (128K + pro voice) → viku-moe (frankenMoE on your 8GB via streaming). Every version is an update to the same HF repo — re-download in LM Studio and you're current.

---

Built with ❤️ by geethudinohtg — the captain runs on your machine. Apache 2.0.