Maritaca AI releases S5-nano 4B model with 262144-token window
Maritaca AI announced its S5-nano model, a 4B-parameter open-weight language model designed for Latin American use. The model supports two modes: Think, which reasons before answering (default), and Instruct, which provides direct responses. It features a 262144-token context window and includes an MTP (multi-token prediction) layer for speculative decoding, optimized for GPUs with native FP8…
Key points
- S5-nano 4B model from Maritaca AI supports 262144-token context window and two modes: Think (reasoning) and Instruct (direct)
- Optimized for FP8 GPUs (Ada, Hopper, Blackwell) with MTP layer for speculative decoding; GGUF weights in BF16, Q8\(0\), Q4\(K\)\(M\)
- Benchmark results compare against LatamGPT 70B and LFM2.5 (24B total, 2B active parameters) across metrics like BLUEX and Prosa-PT
The model’s weights are available in BF16, Q8(0), and Q4(K)(M) quantization formats via the GGUF repository. Benchmark results compare S5-nano against models like LatamGPT 70B and LFM2.5 (24B total, 2B active parameters). Evaluation metrics include BLUEX, Prosa-PT, Arena-Hard v2, and others, with some failures noted in JSON output or judge context limits. The model’s server supports OpenAI API compatibility and includes flags for speculative decoding, tool calling, and mode switching.
maritaca-ai/S5-nano 4B · the best open-weights model in Brazil
huggingface.co · 6 November 2026
Loading the full article…
This text was published by huggingface.co. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- OpenAI launches GPT-6 Cyber, finds 81 Firefox flaws and 948 Linux exploits · 23 src
- US accuses Chinese executive of allegedly sharing Nvidia chip info with Beijing · 11 src
- Moonshot AI releases Kimi K3.1 Pro as open-weights model · 2 src
- OpenAI hints at GPT-6.1 Astra release amid funding and model speculation · 1 src
- NousResearch fine-tunes Qwen3.8-27B into Arete-27B to reduce propaganda in responses · 1 src
Comments
via GitHub Discussions