Complete Guide: The LLM Landscape You Need to Master—Chapter 1

Complete Guide: The LLM Landscape

1. Today's Landscape: Proprietary vs. Open Source

The situation in 2026 is fundamentally different from two years ago. The gap between proprietary and open-source models has narrowed dramatically.

Proprietary Models (GPT-4o, Claude, Gemini) offer the best overall performance, robust APIs with enterprise SLAs, and constant updates without user effort. However, they involve per-token costs that scale with usage, vendor dependence (vendor lock-in), data leaving your infrastructure, and less control over model behavior.

Open-Source Models (LLaMA, Mistral, Qwen, DeepSeek) allow local execution with full data control, predictable costs (fixed infrastructure), customization through fine-tuning, and no rate limits. The trade-off is that they require technical deployment expertise, maintenance is your responsibility, and they generally perform less well on complex reasoning tasks.

The key point for your corporate clients: It is not “one or the other.” The winning strategy is hybrid, using proprietary models for complex or critical tasks and open-source models for high volumes, sensitive data, or specific cases where fine-tuning makes a difference.

2. Transformer Architecture: The Fundamentals

2.1 Tokenization: The First Step

Before a model processes text, it must convert it into numbers. Tokenization splits text into units called tokens.

Take the example “Interbank ofrece préstamos” (“Interbank offers loans”). A model might tokenize it like this: ["Inter", "bank", " ofrece", " prést", "amos"] — 5 tokens.

The main methods are BPE (Byte Pair Encoding), used by GPT and LLaMA, which learns frequent subwords from the training corpus. There is also SentencePiece, used by multilingual models, which treats text as a sequence of bytes, useful for languages without spaces. Finally, WordPiece is the approach used by BERT and some Google models.

Why it matters in practice: A tokenizer trained mainly in English will be inefficient in Spanish, using more tokens for the same content, which means higher costs and less available context. Qwen and multilingual models have an advantage here.

2.2 Embeddings: Meaning as Vectors

Once tokenized, each token is converted into a high-dimensional vector (typically 768, 1024, or 4096 dimensions). These vectors capture semantic relationships.

Imagine a space where “king” − “man” + “woman” ≈ “queen.” That is what embeddings achieve: representing meaning as geometry.

In a Transformer, there are Token Embeddings that represent the lexical meaning of each token, and Positional Embeddings that encode position in the sequence (because attention has no inherent notion of order).

2.3 Attention Mechanism: The Heart of the Transformer

This is the central innovation of the paper “Attention Is All You Need” (2017).

The intuition: When you read “The bank approved the loan because they had good collateral,” how do you know “they” refers to the applicant and not the bank? Your brain connects relevant words even when they are far apart. Attention does exactly that.

The mathematical mechanism works as follows: for each token, the model calculates three vectors. Query (Q) asks “what information do I need?” Key (K) says “this is the information I have.” Value (V) contains “this is the actual content.”

The formula is: Attention(Q,K,V) = softmax(QK^T / √d_k) × V

Breaking it down: QK^T calculates how relevant each token is to every other token. Dividing by √d_k (the square root of the dimension) stabilizes gradients, preventing values from becoming too large. The softmax converts these scores into probabilities that sum to 1. Multiplying by V produces a weighted mix of the relevant information.

Multi-Head Attention runs this process multiple times in parallel (typically 8–96 “heads”), each learning to attend to different kinds of relationships: syntactic, semantic, coreference, and so on.

2.4 The Complete Architecture

A Transformer stacks multiple identical blocks. Each block contains a Multi-Head Attention layer that connects information across positions, followed by a Feed-Forward layer that processes each position independently, with residual connections and normalization that stabilize deep training.

Modern models have 32 to 128+ layers. More layers enable more abstract representations and more complex reasoning, but increase computational cost quadratically.

3. Open-Source Model Families

3.1 LLaMA (Meta)

Origin: Meta AI, beginning with LLaMA 1 in February 2023.

Current versions: LLaMA 3.1 and 3.2 (2024), with 8B, 70B, and 405B parameter variants.

Strengths: Excellent overall performance, extensive documentation, a mature tool ecosystem (llama.cpp, Ollama), and a permissive license for commercial use.

Limitations: The tokenizer is optimized for English, and the larger models require substantial hardware.

Ideal use case: When you need a well-documented general model with broad community support.

3.2 Mistral

Origin: Mistral AI (France), founded by former DeepMind and Meta researchers.

Key models: Mistral 7B, Mixtral 8x7B (MoE architecture), Mistral Large.

Main innovation: Mixtral uses Mixture of Experts (MoE), where not all parameters activate for each token. It has 47B total parameters but activates only ~13B per inference, achieving large-model performance at small-model cost.

Strengths: Exceptional efficiency, good performance in Spanish and European languages, innovative architecture.

Ideal use case: When you need an optimal balance between performance and resources, especially for deployment on limited hardware.

3.3 Qwen (Alibaba)

Origin: Alibaba Cloud, with aggressive development since 2023.

Current versions: Qwen 2.5 with variants from 0.5B to 72B, plus specialized versions (Qwen-Coder, Qwen-Math).

Distinctive strengths: Better multilingual tokenizer (crucial for Spanish), excellent at coding and mathematics, surprisingly capable small models, and an Apache 2.0 license.

The multilingual factor: Qwen processes Spanish with significantly fewer tokens than LLaMA, meaning more effective context and lower costs.

Ideal use case: Spanish or multilingual projects, coding applications, and situations where you need small but capable models.

3.4 DeepSeek

Origin: DeepSeek (China), with massive investment in research.

Key models: DeepSeek-V2, DeepSeek-Coder, DeepSeek-R1.

Innovation: DeepSeek-V2 uses Multi-head Latent Attention (MLA), dramatically reducing KV-cache size, enabling longer contexts with less memory.

DeepSeek-R1 introduced extended chain-of-thought, competing with OpenAI's o1 in mathematical and coding reasoning.

Strengths: Efficient architecture, excellent at reasoning and coding, very competitive API prices.

Consideration: For some corporate clients, its Chinese origin may factor into compliance decisions.

4. Selection Criteria for Use Cases

4.1 Decision Framework

I propose a six-dimensional framework you can use with your clients.

First dimension: Privacy/Compliance Requirements. Can data leave the infrastructure? If not, you need local open-source deployment. Consider sector regulations such as SBS rules for banks.

Second dimension: Task Complexity. For complex reasoning, multistep analysis, or tasks requiring broad knowledge, proprietary models still have an advantage. For specific, well-defined tasks, fine-tuned open-source models can outperform proprietary ones.

Third dimension: Volume and Scale. Low volumes with varied tasks favor proprietary models (pay per use). High volumes with predictable tasks favor open-source models (fixed infrastructure costs are spread across usage).

Fourth dimension: Required Latency. Strict real-time requirements (<100 ms) may require small local models. Batch processing tolerates larger models or APIs.

Fifth dimension: Primary Language. If Spanish predominates, Qwen has a tokenization advantage. For English, LLaMA and Mistral are solid options.

Sixth dimension: Specialization. For coding, consider DeepSeek-Coder or Qwen-Coder. For long-document analysis, prioritize extended-context models.

4.2 Practical Decision Matrix

For customer service chatbots with high volumes and structured responses, the recommendation is a fine-tuned Qwen 7B or Mistral 7B, with fallback to a proprietary model for complex cases.

For legal/financial document analysis where accuracy is critical and sensitive data is involved, the option is local deployment of LLaMA 70B or Qwen 72B with robust RAG.

For marketing content generation requiring creativity and variety, proprietary models such as Claude or GPT-4o are preferable, since the cost per piece is low and quality matters more.

For internal coding assistants using proprietary company data, local DeepSeek-Coder or Qwen-Coder models fine-tuned on the internal codebase are the best option.

For data extraction from invoices/documents with high volumes and predictable structures, small models (Qwen 1.5B, Mistral 7B) with targeted fine-tuning work perfectly.

4.3 The Hybrid Approach in Practice

For a bank such as Interbank, a realistic architecture would be: a local open-source model (Qwen 14B) as the first line for frequent queries and sensitive data, intelligent routing that escalates to a proprietary model (Claude/GPT-4) for complex or ambiguous cases, and specialized fine-tuned models for specific tasks such as transaction classification or fraud detection.

Key Points for Your Training Sessions

On Transformers: The magic is in attention, which allows the model to “look at” the entire context simultaneously, unlike earlier architectures that processed it sequentially. This enables an understanding of long-range dependencies.

On Open Source: It is no longer the “poor relation.” For many business use cases, models such as Qwen 14B or Mistral 8x7B are sufficient and offer control and cost advantages proprietary models cannot match.

On Selection: No model is universally best. The right question is not “what is the best model?” but “what is the best model for this specific use case, with these data, budget, and latency constraints?”

Click to read the: Next Chapter