My Journey

AI models galore

7 min read

Light / Background / Routine Tasks

  • comparison
  • Purpose: Fast, low-resource, suitable for simple queries, summarization, classification, and background automation.
  • Top 3 Models:
    1. Minstral 3.1 3B (OpenRouter) — your current choice; very fast, low latency
      1. includes image input
    2. meta-llama/llama-3.2-3b-instruct (Meta)
    3. Phi 4 (Microsoft)

How “Light Reasoning” Operates in the Humanities

In STEM, “reasoning” usually refers to deterministic logic—deductive proofs, execution steps, or mathematical steps. In contrast, humanities and social science reasoning relies on abductive and inductive logic:

  • Contextual synthesis: Connecting historical events, sociological theories, or philosophical concepts across eras.
  • Nuance and trade-offs: Balancing opposing perspectives (e.g., assessing economic policies or ethical frameworks) without a single “correct” answer.
  • Argumentative structure: Building a coherent thesis, managing transitions, and evaluating qualitative evidence.

NOTE

all AI models developed and deployed in China, is subject to China’s censorship and content regulation rules. These include:

  • Content safety filters that block politically sensitive material, misinformation, and content deemed harmful to national security or social stability.
  • Compliance with the “Generative AI Service Management Interim Measures” issued in 2023 by China’s Cyberspace Administration, which requires all AI services to ensure content aligns with socialist core values and national regulations.
  • Mandatory real-name registration and data localization for users and service providers.

These rules apply broadly to all major Chinese AI models, including Baidu’s ERNIE Bot, Alibaba’s Qwen, Tongyi Lab’s Qwen, and Moonshot AI’s Kimi.

Performance of These Specific Models in Qualitative Tasks

+-------------------------------------------------------------------------------+
| MODEL COMPARISON FOR HUMANITIES & SOCIAL SCIENCES                             |
+-------------------------------------------------------------------------------+
| Model             | Size   | Knowledge Retention | Nuance & Prose Quality   |
+-------------------+--------+---------------------+--------------------------+
| Phi-4             | 14B    | Highest             | Analytical, highly formal|
| Llama 3.1 8B      | 8B     | High                | Balanced, natural tone   |
| Ministral 3B      | 3B     | Moderate            | Direct, concise          |
| Llama 3.2 3B      | 3B     | Moderate            | Fast, basic structure    |
+-------------------------------------------------------------------------------+

1. Phi-4 (14B)

  • Strengths: At 14B parameters, Phi-4 has significantly more world knowledge stored in its weights than the 3B or 8B models. It excels at multi-perspectival synthesis—such as comparing two philosophical movements or analyzing cause-and-effect in historical events.
  • Limitations: Its pre-training heavily emphasizes formal logic and synthetic instruction data. As a result, its prose style can occasionally feel somewhat rigid or overly structured for qualitative essay writing.

2. Llama 3.1 (8B Instruct)

  • Strengths: Widely considered the “sweet spot” for standard qualitative tasks among small models. Meta tuned Llama 3.1 extensively on conversational and instruction data. It handles qualitative reasoning—like critiquing an essay, generating counterarguments, or summarizing social theory—with fluid, natural phrasing.
  • Limitations: Lacks the deep niche historical recall of 70B+ or frontier models, so it may occasionally generalise specific historical dates or lesser-known sociologists if unprompted.

3. Llama 3.2 (3B) & Ministral (3B)

  • Strengths: Exceptional speed and low resource usage. Great for first-pass outlining or summarizing short texts.
  • Limitations: At 3 billion parameters, parameter budget restricts broad humanities knowledge. When asked for deep qualitative analysis, 3B models tend to produce surface-level bullet points or rely on generic summaries rather than deep contextual synthesis.

Unlocking Qualitative Reasoning via Prompting

Because these non-reasoning models lack internal “thinking tokens,” you can unlock their latent reasoning capabilities by guiding their output structure manually:

  • Use Explicit Chain-of-Thought: Ask the model to evaluate opposing arguments before stating a conclusion.
  • Assign a Specific Perspective: Prompting with “Critique this policy through a functionalist sociological lens” gives the model a concrete framework to evaluate.

Content Relevant to You Personally

1. Document Analysis and Obsidian Integration

For organizing thoughts, notes, or essays in Obsidian:

  • Llama 3.1 8B and Phi-4 are ideal for summarizing long articles, generating literature outlines, or extracting core themes from social science texts.
  • Standard Instruct models yield clean Markdown immediately—allowing you to copy structural outputs straight into your notes without manually stripping tags.

2. Crafting Custom Prompts for Qualitative Tasks

If you want to use these standard models for deeper humanities exploration, standardizing a structured prompt template yields strong results:

Role: Social Sciences Assistant

Task: Analyze the provided text/topic.

Instructions:

  1. Identify 2-3 core theoretical themes.
  2. Outline the primary arguments for each theme.
  3. Present at least one common critique or counter-perspective.
  4. Provide a structured summary in clean Markdown.

Relevant and General Content

For qualitative tasks like social sciences, humanities, policy analysis, and document synthesis, small 3B–14B models can sometimes lack the deep cultural and historical context required for subtle arguments. However, several ultra-low-cost (and even free) models on OpenRouter bridge this gap by offering high-parameter knowledge depth at a fraction of a cent per request.

Top Low-Cost Models for Humanities & Social Sciences

+-------------------------------------------------------------------------------+ | COST VS. QUALITATIVE REASONING MATRIX | +-------------------------------------------------------------------------------+ | Model | Params | Typical OpenRouter Price | Best Qualitative Use | +---------------------+--------+--------------------------+---------------------+ | Llama 3.3 70B | 70B | ~0.35 / M tokens | Deep analysis, essay| | Qwen 2.5 / 3.6 32B+ | 32–72B | ~0.30 / M tokens | Synthesis, theories | | Gemma 2 / 4 27B–31B | 27–31B | ~0.15 / M tokens | Editorial, critique | | DeepSeek V3 / Flash | MoE | ~0.28 / M tokens | High-volume research| +-------------------------------------------------------------------------------+

1. Meta: Llama 3.3 70B Instruct

  • Price Tier: Ultra-cheap (~$0.10 per million input tokens on endpoints like DeepInfra or Novita; frequently available via :free endpoints on OpenRouter).
  • Why it shines in Humanities: At 70 billion parameters, it holds immense broad-world knowledge across history, sociology, and political theory. It evaluates arguments with nuance, balances opposing viewpoints smoothly, and produces natural, articulate prose without sounding like a robotic textbook.

2. Alibaba: Qwen 2.5 / 3.6 (32B or 72B)

  • Price Tier: Ultra-cheap (~0.20 per million tokens).
  • Why it shines in Humanities: Qwen is widely considered one of the strongest open-weights families for non-STEM subjects. It excels at cross-referencing ideas, analyzing non-Western and global history, and handling subtle distinctions in policy or philosophy.

3. Google: Gemma 2 27B / Gemma 4 31B

  • Price Tier: Very low cost or free on OpenRouter.
  • Why it shines in Humanities: Google’s Gemma models are trained heavily on high-quality editorial and academic text. As a result, their writing output is refined, making them ideal for critiquing written work, summarizing social science literature, and stylistic polishing.

4. DeepSeek V3 / V4 Flash (Non-Reasoning Instruct)

  • Price Tier: Extremely low (~0.28 output per million tokens).
  • Why it shines in Humanities: While DeepSeek R1 gets attention for hard STEM math, standard DeepSeek V3 / Flash is a fast, standard instruct model. Its Mixture-of-Experts (MoE) architecture makes it remarkably cheap for processing large blocks of qualitative text, extracting core arguments, and comparing sociological frameworks.

OpenRouter Cost-Saving Hacks

To keep your API costs near zero while testing these models:

  • Append :floor to Model Slugs: Requesting meta-llama/llama-3.3-70b-instruct:floor automatically routes your prompt to the single cheapest provider serving that model at that moment.
  • Use OpenRouter :free Variants: OpenRouter hosts free endpoints for several larger models (e.g., llama-3.3-70b-instruct:free or gemma-4-31b-it:free). These are ideal for draft generation and casual exploratory reading.

Content Relevant to You Personally

1. Seamless Obsidian Integration

Larger standard models like Llama 3.3 70B and Qwen 2.5 32B/72B generate exceptionally clean Markdown headers and lists natively.

  • They adhere closely to formatting rules, making their outputs easy to record directly into your Obsidian notes or canvas files without extra cleanup.
  • They avoid inserting arbitrary internal thinking steps, ensuring your notes remain structured, readable, and ready for future reference.

2. Conceptual Exploration & Synthesis

When exploring non-technical topics—such as historical timelines, social policy comparisons, or literature summaries:

  • Llama 3.3 70B offers the best balance of rich background knowledge and fluid, articulate writing.
  • Qwen 2.5 32B or 72B provides exceptional analytical depth for comparing complex philosophical or social frameworks at a fraction of the cost of proprietary frontier models.

Moderate / Daily Reasoning & Instruction Following

  • Purpose: Good balance of speed and reliability, suitable for instruction-following, basic reasoning, and moderate creative tasks. Include Image inputs (screenshots etc)
  • Comparison

Advanced Reasoning / Complex Tasks

  • Purpose: Deep reasoning, multi-step problem solving, coding, and nuanced understanding.

High-Performance / Specialized Tasks

  • Purpose: Scientific research, large-scale data analysis, multi-modal inputs, or high-stakes decision-making.
  • Top 3 Models:
    1. GPT-5 (if accessible) — cutting-edge, large context, multi-modal.
    2. OpenAI GPT-4.1 Pro — for maximum accuracy and reasoning depth.

Summary

  • You’re currently comfortable with Minstral 3B for routine/light tasks.

Graph View

Backlinks