Is S5-nano state of the art?Pending benchmarks.Announced Nov 2, 2026; no independent results yet.
Maritaca AI's open-weight ~4B-parameter (≈4.2B text; HF safetensors list ~5B including vision) Portuguese/English multimodal model, continued from Qwen3.5-4B. Think (default) and Instruct modes, 262,144-token context, and an MTP layer for speculative decoding. Weights on Hugging Face as BF16, FP8 and GGUF under Apache-2.0 for self-hosting with vLLM, SGLang or llama.cpp; no Maritaca-hosted API for this model.
The details so far
- Lab
- Maritaca AI
- Announced
- Nov 2, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- ~4B text (≈4.2B; HF safetensors ~5B incl. vision)
- Role
- Open-weight chat / reasoning model with Think and Instruct modes; image-text-to-text
- Base
- Continued pretrain from Qwen3.5-4B; separate from the closed S5 flagship and from the Qwen3.5-4B catalog page
- Context
- 262,144 tokens
- Languages
- Portuguese and English (HF card)
- Checkpoints
- maritaca-ai/S5-nano (BF16), maritaca-ai/S5-nano-FP8, maritaca-ai/S5-nano-GGUF (BF16 / Q8_0 / Q4_K_M + mmproj)
- Serving
- Self-host: vLLM, SGLang, llama.cpp. No Maritaca-hosted API stated for S5-nano
- License
- Apache-2.0
- Evals hub
- Observatório Maritaca (github.com/maritaca-ai/observatorio-maritaca)
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: on Maritaca's own Portuguese open-model comparison (10 models; Observatório Maritaca / model-card Totals table), S5-nano Think total average is 75.25 with 23/23 coverage, above the next listed open models in that table (e.g. Qwen3.5-4B Think 68.68).
Lab claim: S5-nano Instruct total average 66.51 (23/23) on the same Maritaca comparison table.
Lab claim: Native BR Think average 83.74 and Instruct 74.65 on that table.
Sources
REPOmaritaca-ai/S5-nano model card (Hugging Face)huggingface.co/maritaca-ai/S5-nanoREPOObservatório Maritaca (GitHub)github.com/maritaca-ai/observatorio-maritacaLABMaritaca AI on X: S5-nanox.com/MaritacaAI/status/2118781394524463564 · Nov 6, 2026LABMaritaca AI on X: S5-nano weightsx.com/MaritacaAI/status/2118768446225920029 · Nov 6, 2026Tracking continues until an independent board lists it.