Local AI is Finally Ready: 32GB Intel GPU + Qwen 3.8

Local AI on the Intel B70: A Practical Sweet Spot for On-Premise Workloads

In this video, tech enthusiast Lon Cyon shares his hands-on experience running the Qwen 3.8 dense model locally on an Intel B70 32GB GPU (ASRock model). He positions the card as the most affordable 32GB GPU on the market by 2026 standards and demonstrates how it handles real business tasks—legal document summarization, coding, RSS filtering, and web scraping—entirely on local hardware, with no cloud tokens consumed.

Hardware & Setup

• GPU: Intel B70 with 32GB VRAM, currently selling for about $1,300 (it was under $1,000 a few weeks ago). Lon notes that a lower-tier B60/B65 model has the same memory but less compute power.
• Host system: A Minisforum mini PC with an AMD Strix processor, connected via an Oculink GT box. The GPU can also work in any PC with a PCI Express slot.
• Software: Ubuntu 26.04, Qwen 3.8 (a 26-billion-parameter dense model, not a mixture of experts), and command-line coding harnesses like OpenCode and Codex.
• Disclosure: The GPU was self-funded; the Oculink box and mini PC were provided free by manufacturers. No other compensation or approval was involved.

Performance & Power

• Token generation: About 25 tokens per second with heavy context loaded, rarely dropping below 20. Slower than Nvidia cards, but “fast enough” for practical work.
• Context capacity: The model plus roughly 96,000 tokens of context fit entirely on the GPU, avoiding CPU/RAM fallback.
• Power draw: Approximately 322 watts under load and 55 watts at idle with the model loaded. The fan is loud, and Lon warns that this will show up on your electric bill.

Real-World Use Cases

• Legal summarization: Loaded a DJI court case and generated a clear summary with a chart showing which points DJI won and lost—all locally.
• Coding projects: Built a Space Invaders clone with the best visuals of any local model he tested, and a Gadget Tracker app that scans RSS feeds every 15 minutes, uses Qwen to filter gadget articles, and stores results in a SQLite database.
• Automated web scraping: Extracted NYC Council member names and email addresses from HTML source code, generated a verified CSV in under two minutes. A similar PDF-based contact list task also worked flawlessly.

Workflow & Optimization

• Initial performance was slow, but recent software updates brought significant gains.
• Lon uses AI assistants like OpenAI’s Codex as development partners to configure Linux, and asks the client to output a report of all changes for future reference.
• OpenCode periodically compacts context to manage the 96k-token limit, keeping essential information while discarding unnecessary details.
• The model spends more time in reasoning mode, checking itself before coding, which produces many tokens but results in bug-free, one-shot wins more often.

Limitations & Considerations

• Dense 26B models require patience: reasoning tokens are generated before any code appears.
• 32GB is described as the sweet spot for local dense models; lower-memory cards require more frequent context compaction.
• Idle power draw prevents 24/7 operation, and solving the power problem is the next step.
• Nvidia alternatives are faster but cost significantly more; Lon suggests a Spark mini PC as a budget alternative.

Key Takeaway

The Intel B70 + Qwen 3.8 combination delivers roughly 80% of what Lon typically uses Gemini, Claude, or ChatGPT for—completely locally. It is not the fastest or most power-efficient option, but for users who value data privacy, no token costs, and solid performance on a modest budget, this setup is a compelling consumer sweet spot.