Scientific Reasoning

1. Install Ollama → ollama pull deepseek-r1:32b (20 GB) or ollama pull qwen-3-30b-a3b (18 GB — MoE, strong reasoning).
2. For physics problems: ollama run deepseek-r1:32b → "A 2 kg block slides down a frictionless 30° incline. Calculate the acceleration and the time to slide 5 meters. Show your work step by step."
3. The reasoning model outputs its chain-of-thought (hidden by default) then the answer. First response in 10-30 seconds on 24 GB GPU.
4. For multi-step scientific reasoning (design an experiment, analyze results): use the same reasoning models. Prompt: "Design an experiment to test whether a new fertilizer increases plant growth. Include control group, sample size, statistical test, and potential confounding variables."
5. For domain-specific science (quantum mechanics, relativity, molecular biology): reasoning models handle the logical/mathematical aspects but lack deep domain knowledge. Supplement with RAG over textbooks for niche topics.
6. Evaluation: pip install lm-evaluation-harness → test on GPQA, MMLU-Pro, ARC-Challenge to benchmark your local model against published results.

Used RTX 3060 12 GB (~$200-250, see /hardware/rtx-3060-12gb). Runs DeepSeek R1 Distill Llama 8B at 50-80 tok/s or Qwen 7B distill at 40-60 tok/s. These handle high-school to intro-college physics, chemistry, and biology problems competently (GPQA ~30-40%). For undergraduate-level scientific reasoning: the 14B distilled models (Qwen 14B) run at 25-35 tok/s with noticeably better multi-step reasoning. Pair with Ryzen 5 5600 + 32 GB DDR4 + 512 GB NVMe. Total: ~$400-480. $400 gets you competent undergrad science reasoning; graduate-level requires 32B+ models.

Used RTX 3090 24 GB ($700-900, see /hardware/rtx-3090). Runs DeepSeek R1 Distill Qwen 32B at 15-25 tok/s — handles graduate-level physics and chemistry problems (GPQA ~50-65%). For research-grade scientific reasoning: Qwen 3 235B MoE IQ4_XS (50 GB) on dual RTX 3090 (48 GB total, ~$1,600) at 5-10 tok/s — GPQA 70%+, near-frontier quality. Total: ~$1,800-2,500. Scientific reasoning benefits disproportionately from model scale — the jump from 7B to 32B to 235B is qualitative, not just quantitative. Each step unlocks a new tier of scientific problems.

The mistake: Using a non-reasoning chat model for scientific problem-solving, getting a confidently wrong answer, and citing it in a paper or homework. Why it fails: Standard LLMs don't do step-by-step verification. Asked "What's the pH of 0.1M HCl?" a chat model might say "pH = 1" (correct) or "pH = 0.1" (confusing concentration with pH) or "pH = 13" (confusing acid with base) — all with equal confidence. Without a reasoning trace, you can't tell which answers were reasoned and which were hallucinated. The fix: Use a model with explicit chain-of-thought reasoning (DeepSeek R1 distillation, Qwen 3 with thinking mode). These models output their reasoning before the answer. Read the reasoning — if the logic is garbage, the answer is garbage. Also: verify calculations independently (Wolfram Alpha, Python). The model is a reasoning partner, not a calculator — it makes arithmetic errors even when the logic is correct. Trust the reasoning trace, verify the numbers.