Small Language Models: The 2026 Landscape
Created by Adrian Dunkley | maestroailabs.com | ceo@maestrosai.com | Fair Use
This is a selection guide, not a ranking or a complete catalogue. Primary model documentation checked: 11 November 2026. Models below are documented examples worth evaluating for a defined task. Their inclusion does not establish that they are the newest, fastest or best for a Caribbean or Latin American language.
Record the exact checkpoint and runtime version you test. Similar family names can refer to different sizes, modalities, licenses and training methods.
Documented examples
Parameter counts below describe model architecture, not download size or required RAM. “B” means billion parameters.
| Exact family or checkpoint | Published sizes relevant to this guide | License / terms | Important distinction |
|---|---|---|---|
| Microsoft Phi-4-mini-instruct | 3.8B | MIT | Text model; evaluate each working language |
| Microsoft phi-4 | 14B | MIT | Text input, primarily English-oriented; not an OCR engine |
| Google Gemma 3 | 270M, 1B, 4B, 12B, 27B | Gemma terms | 270M/1B are text-only; larger listed variants support images |
| Google Gemma 4 | E2B, E4B, 12B, 26B A4B, 31B | Apache 2.0 | Effective and active counts are not total stored weights |
| Qwen3 original release | Dense: 0.6B, 1.7B, 4B, 8B, 14B, 32B | Apache 2.0 | This table names the original dense release, not every Qwen variant |
| Mistral-NeMo-Instruct-2407 | 12B | Apache 2.0 | Text model; do not apply its license or size to all Mistral models |
The earlier Gemma 3/4 “9B” and Qwen3 “7B” recommendations are not the documented variants listed here. Do not guess a runtime tag from a family name. Choose an actual model-card entry and confirm that your runtime supports its architecture and file format.
How to compare candidates
Text drafting and structured extraction
Compare a small text model such as Phi-4-mini-instruct or Qwen3-4B against a second model on the same material. Treat these as candidates, not recommendations backed by a regional benchmark. Measure missing fields, invented details, invalid structure and the amount of correction required.
For scanned invoices, use a separately tested OCR or vision component. A text-only model cannot inspect a receipt image simply because the task is called “document processing.” Recalculate totals with code and keep human approval before accounting entries or payments.
Multimodal input
Match the exact variant to the input type. Gemma 3’s smaller text variants do not have the same input capabilities as its larger variants. The current Gemma 4 card lists image input across the family and audio input for E2B, E4B and 12B. The runtime must support the corresponding inputs too. Gemma 3 card; Gemma 4 card.
Image support does not establish reliable handwriting recognition, diagnosis or document authenticity. Build those checks into the task evaluation.
Regional language work
A multilingual training claim is not a local-language quality score. Test code-switching, currency notation, place names, customer tone and preservation of meaning. A French benchmark is not evidence for Haitian Kreyòl, and Spanish performance is not evidence for Quechua.
Have competent speakers evaluate unfamiliar languages and important customer communications. Do not use a model’s self-rated confidence as the sole rule for deciding when a person is needed.
Why active parameters can mislead
Meta’s Llama 4 Scout model card lists approximately 109B total parameters, although the model uses about 17B active parameters per token. Its advertised long context does not make it a small-laptop default.
As a rough arithmetic illustration, 109 billion weights at exactly 4 bits would occupy 54.5 GB before quantization metadata, higher-precision components, attention cache and runtime overhead. That alone exceeds 24 GB of GPU memory. CPU offloading or other specialized configurations change the deployment tradeoff; they do not make the full model fit in a 24 GB GPU. Do not buy hardware on the assumption that “17B active” means “17B stored.”
Gemma 4’s E2B and E4B labels also describe effective counts. Consult the model card’s full parameter breakdown, including embeddings and modality components, before estimating memory.
Estimate memory, then measure it
For a dense model, a simplified weight-only estimate is:
weight bytes ≈ parameter count × bits per weight ÷ 8
| Hypothetical dense model | At 16 bits | At 8 bits | At 4 bits |
|---|---|---|---|
| 4B parameters | 8 GB | 4 GB | 2 GB |
| 8B parameters | 16 GB | 8 GB | 4 GB |
| 14B parameters | 28 GB | 14 GB | 7 GB |
| 32B parameters | 64 GB | 32 GB | 16 GB |
These are decimal-GB arithmetic examples, not hardware requirements or actual file sizes. Quantization schemes do not store every component at exactly the named bit depth. Add the operating system, runtime buffers, attention cache, modality components and concurrency. GPU memory and system RAM are not interchangeable at the same speed.
A longer context or several simultaneous requests can make a previously working setup run out of memory. Test the workload you intend to use, not just whether the model can answer “hello.”
Licensing and provenance
- Open weights means weights are available; it does not mean every use is unrestricted.
- Read the exact checkpoint’s license, any acceptable-use terms and redistribution requirements. Preserve required notices.
- A community quantization or fine-tune has its own provenance. Check its base model, publisher and changes.
- Training-data rights, privacy duties and application security are separate from permission to run model weights.
- Do not download a model package that requires arbitrary remote code execution without technical review. Prefer supported formats from an identifiable publisher.
Do not reject a working model solely because it is older, or adopt a new model based only on a leaderboard. Maintenance, runtime support and measured task quality matter.
A model-selection record
Keep a short record for each candidate:
Publisher, checkpoint, revision/digest and license:
Runtime and version; quantization/file format:
Hardware; context limit; concurrency:
Authorized data and offline/cloud mode:
Test cases, languages and reviewer:
Errors by severity; accepted results; correction time:
Latency, peak memory and cost per accepted result:
Known limitations and manual fallback:
Decision, owner and review date:
If a source does not provide a model version, test setup and reproducible evidence, do not turn its claim into a numerical ranking. See deployment, use cases and the rankings hub for related guidance.
Created by Adrian Dunkley | MaestrosAI | maestroailabs.com | ceo@maestrosai.com Fair Use, Educational Resource | April 2026