​

Small Language Models: The 2026 Landscape

Created by Adrian Dunkley | maestrosai.com | ceo@maestrosai.com | Fair Use

​

The 2026 SLM short list

​

Deep dive on the five most useful for LAC SMBs

​

Microsoft Phi-4 and Phi-4 Mini

• Why it matters for LAC: Phi-4 punches well above its weight on reasoning and math. It’s excellent for invoice extraction, accounting reconciliation, and any structured-data task. Phi-4 Mini at 3.8B runs comfortably on an older laptop.
• Weakness: English-first. For Spanish and Portuguese output it works, but you’ll want to fine-tune on a few hundred examples of your business’s style.
• Licensing: MIT. You can run it commercially, modify it, ship it.
• Hardware: Phi-4 Mini (3.8B) runs on 8 GB RAM. Phi-4 (14B) runs on 16-24 GB.
• Good for: Back-office tasks, document processing, offline assistants.

​

Google Gemma 4

• Why it matters for LAC: Gemma 4 (April 2026) is purpose-built for reasoning and agentic workflows. The 9B model hits a sweet spot: strong output, modest hardware, broad multilingual coverage, and it fine-tunes cleanly.
• Multilingual quality: 140+ languages in training data, with Brazilian Portuguese, Mexican Spanish, Argentine Spanish, and French handled well out of the box. Kreyòl and Papiamento need fine-tuning or human review.
• Hardware: 2B on 6 GB RAM, 9B on 16 GB, 27B on 32+ GB.
• Good for: Customer-service agents, content drafting, summarisation, the brain of a privacy-first WhatsApp agent.

​

Meta Llama 4 Scout

• Why it matters for LAC: At 17B active parameters (MoE with 109B total), Scout is the largest practical “small” model in 2026. It supports a 10-million-token context window, which means it can hold a whole year of business documents in memory. Multimodal, too.
• Multilingual: pre-trained on 200 languages including over 100 with >1 billion tokens each. Top-tier Spanish and Portuguese. Good French and Dutch (helpful for Curaçao, Aruba).
• Hardware: full-precision Scout needs a workstation GPU. Quantised Q4 runs on a 24 GB RTX 4090 or an Apple Silicon machine with 64 GB unified memory. Commercial Mac Studios are a realistic target.
• Good for: Privacy-first agents that need long context (legal review, multi-week case histories, large knowledge bases).

​

Mistral 7B / NeMo / Small

• Why it matters for LAC: Apache 2.0 on most variants (no gotchas). Strong in Romance languages, including French Caribbean content. Easiest family to fine-tune with modest data.
• Multilingual: French, Spanish, Italian, Portuguese, and English are all strong. NeMo in particular is trained with multilingual emphasis.
• Hardware: 7B Q4 runs on 8 GB RAM. NeMo 12B wants 16 GB.
• Good for: French Caribbean use cases (Martinique, Guadeloupe, Haiti), European Spanish/Portuguese, bilingual content.

​

Qwen 3

• Why it matters for LAC: Extremely strong reasoning for its size, very fast inference, flexible size options from 0.5B to 32B. Apache 2.0 licensed. The 7B and 14B variants are standouts for mid-range hardware.
• Multilingual: strongest in English and Chinese, competent in Spanish and Portuguese, weaker in French Caribbean and Kreyòl (review required).
• Hardware: Qwen 3 7B runs on 8 GB, 14B on 16 GB, 32B on 24 GB (Q4).
• Good for: High-throughput tasks where latency matters: voice assistants, POS integrations, real-time dashboards.

​

Scorecard: picking the right SLM for a task

​

What to avoid

• Models older than 12 months unless they have a clear niche. The field is moving fast; 2024-era SLMs are rarely worth setting up in 2026.
• Models without open weights if your goal is offline deployment. You can’t run them locally.
• Licenses with usage restrictions that your business model might trip. Read the license for Llama 4 and Gemma before shipping a product.
• Unreviewed Kreyòl, Papiamento, or indigenous-language output. No 2026 SLM is reliable here without human review.

​

Trends for the next 6 months

• Multimodal SLMs are catching up to cloud models. Expect Gemma 4 and Phi-5 to handle images, charts, and simple video by year-end.
• MoE (mixture-of-experts) SLMs like Llama 4 Scout give frontier-like behavior with consumer hardware. More vendors will ship MoE small models.
• Tool-use SLMs: open-weight models fine-tuned specifically for agent tool use. Watch for Gemma 4 agentic variants and Qwen 3 function-calling releases.
• LAC language fine-tunes: community fine-tunes for Brazilian Portuguese, Caribbean Spanish, and Kreyòl are beginning to emerge on Hugging Face. Track for your market.

​

Related reading

• use-cases.md: concrete LAC scenarios using the models above.
• deployment.md: how to actually install and run.
• rankings/global-benchmarks.md: numeric benchmark scores.
• governance/README.md: SLMs as a data-residency solution.