Local AI in 41 seconds, cloud AI in 21, on the same question
A simple timed experiment: the same question, two Hermes agents, low reasoning effort — 41 seconds locally, 21 seconds in the cloud.
A simple experiment was run, without benchmarks, synthetic workloads or elaborate measurements. Two Hermes agents received the same question. The first was connected to local AI at low reasoning effort. The second was connected to Claude Haiku 5.5 via AWS Bedrock, also at low effort.
Results: local AI, 41 seconds; cloud AI, 21 seconds.
Local hardware, as stated in the post: a laptop costing approximately $4,000–4,500 (€4,000–4,500); NVIDIA RTX 5090 Laptop, 24 GB VRAM; 64 GB of RAM; inference engine Basalt (based on Strata); model Qwen 3.8 Next 3_S; agent framework Hermes; reasoning effort low. On the local demonstration screen the model line reads “Qwen3.8 Flash Next Iq3 S Basalt · Low”. The timed recordings show this prompt:
“Compare the stock market performance of Microsoft and Intel. Provide a detailed analysis in approximately 1,000 tokens.”
The two answers were then given to Google Gemini 4 Argon for a comparison. The scores and commentary below are quoted verbatim from that comparison as published in the thread.
I analyzed two AI responses comparing MSFT and INTC stock performance:
☁️ Response 1: Cloud AI — 73.5 / 100
🔒 Response 2: Local AI — 89.5 / 100
10-Parameter Scorecard (Cloud vs. Local):
1. Data Granularity: 7/10 vs 9/10
2. Time Horizons: 6/10 vs 10/10
3. Earnings Depth: 9/10 vs 8/10
4. Valuation & Margins: 6/10 vs 9/10
5. Risk & Volatility: 7/10 vs 9/10
6. Catalysts & Narrative: 8/10 vs 9/10
7. S&P 500 Benchmarking: 7/10 vs 9/10
8. Structure & Readability: 8/10 vs 9/10
9. Analytical Synthesis: 7/10 vs 9.5/10
10. Factual Consistency: 8.5/10 vs 8/10
Depth Analysis:
• Cloud AI (R1) reads like a standard news wire. It nailed Q1 FY27 earnings specifics ($677B Azure backlog, +48% growth, Israel fab accident) but missed the bigger financial picture.
• Local AI (R2) thinks like a portfolio manager. It exposed why Intel's "+57% YTD" is a statistical illusion after a -59% crash from $142, factored in P/E multiples (28.5x vs negative TTM), highlighted $1.9B quarterly foundry losses on $108B capex, and noted MSFT lagging the S&P 500 over 1Y/3Y windows.
Key Errors & Omissions:
• Cloud AI: Complete valuation blindness — zero mention of Market Cap, P/E ratios, 52-week ranges, or 5-year returns.
• Local AI: One math contradiction — claiming MSFT had a "smooth upward grind" while citing a $349 low in June 2026 (which would mean a ~28% mid-year crash), plus treating a routine law-firm press release as a real risk factor.
Local AI delivered institutional-grade synthesis where the Cloud stopped at surface-level summary.
The same thread describes a practical direction: one interface that uses local AI for private documents and cloud AI for ordinary, time-critical questions, with setup that works without manual API keys. Local and cloud agents could collaborate so that a local agent analyses private documents, forms safe questions for a stronger cloud model, filters sensitive information and combines the answers without exposing private data. The work continues on inexpensive hardware: an existing gaming PC, or an upgrade of about €200–300 to a computer already in an office or at home.
A closing, subjective note in the thread is that more agents, more different models, and more communication between them in a single interface would be useful. That is the MixAI direction already described on this site.