whatmodel.sh / S5-nano
S5-nano
↗ share
+ contribute
S5-nano is an open-weights language model from Maritaca AI, released November 2026. Its published size is 4.66B parameters and a 262k-token context window. Licence: apache-2.0.
Catalogue identity notes
Maritaca AI's Portuguese/English model built on Qwen/Qwen3.5-4B by continued pretraining plus post-training; Think (default) and Instruct modes; text+image (vision encoder) with an MTP layer. BF16 weights committed 2026-11-02 (commit fe2c063); later commits to the pinned sha fe3a358 (2026-11-10) are card edits. LICENSE is Apache 2.0 with 'Copyright 2026 Alibaba Cloud' (inherited Qwen notice). Hosted on Maritaca's own API as S5-nano (256K context there). Results use the Think table (highest reasoning setting); Instruct-mode values not imported. Card scores are Maritaca's own runs of all listed models; other models' columns not used. API price: Maritaca's own model docs (observed 2026-11-12): 'R$ 0.10 per 1M input tokens and R$ 0.60 per 1M output tokens'; accepted API names S5-nano / s5-nano. Prices are in BRL, so no USD price_in/price_out recorded (no conversion); a first-party-maritaca ledger entry would be needed before recording any price.
Model information sources
Release date ↗ · Weights licence ↗ · Parameter count ↗
Dense; HF safetensors metadata total 4,659,865,088 parameters (BF16, includes vision encoder and MTP layer), same convention as qwen3-5-4b (4.66). Card calls it 'a 4B-parameter model' and its parameter table gives 'S5-nano ≈4.2B · text'.
$ how it compares
Scored LLMs released closest to S5-nano, out of 1206 with a score.
| # | model | score | where |
|---|---|---|---|
| 1 | 89±1 | api $16.00/M | |
| 14 | 79±2 | api $3.00/M | |
| 22 | 78±1 | api price unknown | |
| 32 | 76±1 | api $0.20/M | |
| 37 | 75±1 | api price unknown |
$ what it's good at
other tests2 not in a task · 0.0% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LongBench v2 lab | 49.7% ✎ | ▂·▁·▂▂▂▃▃▃█▃ beats 18/35 | — |
| IFBench prompt-level lab | 42.7% ✎ | ▄······██▄▄▄ beats 0/7 | — |
$ where to run it
Through an API
api · price unknown · 262k context · docs.maritaca.ai ↗
Memory is what a version takes in GPU memory (or a Mac's unified memory) with an 8k-token conversation. With less, it runs partly from system RAM, more slowly. Enter your machine to see what runs at full speed.
llama.cpp · recommended
version
- Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.macOS / Linux
curl -LsSf https://llama.app/install.sh | shWindowsirm https://llama.app/install.ps1 | iex - Run S5-nano · needs 5.1 GB memory, 3.46 GB download · from huggingface.co/maritaca-ai (official)chat in the terminal
llama cli -hf maritaca-ai/S5-nano-GGUF:Q4_K_M - Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf maritaca-ai/S5-nano-GGUF:Q4_K_M
LM Studio
version
- Install LM Studio · Open the app once before using the lms command.
- Run S5-nano · needs 5.1 GB memory, 3.46 GB download · from huggingface.co/maritaca-ai (official)
- Or download it in the terminal
lms get https://huggingface.co/maritaca-ai/S5-nano-GGUF@Q4_K_M -y - Then run lms ls to see its name, and lms chat followed by that name
Ollama
version
- Install Ollama
download it (macOS / Windows) ↗
Linuxcurl -fsSL https://ollama.com/install.sh | sh - Run S5-nano · needs 5.1 GB memory, 3.46 GB download · from huggingface.co/maritaca-ai (official)chat in the terminal
ollama run hf.co/maritaca-ai/S5-nano-GGUF:Q4_K_M
Other ways to run it
version
- Run it as published · needs 11.3 GB memory, 9.32 GB download, NVIDIA only · from huggingface.co/maritaca-ai (official)
vllm serve maritaca-ai/S5-nano --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
Model repository and documentation: huggingface.co/maritaca-ai ↗
$ published results · 2 results, not enough overlap for a score
| test | result | source and conditions |
|---|---|---|
| IFBench prompt-level | 42.7% | maritaca-ai/S5-nano model card · maker-reported Maritaca own evaluation, pinned S5-nano card, 'Results by benchmark' table 'Think / Thinking', column 'S5-nano'. Card: 'Results for the original S5-nano checkpoint, evaluated in BF16 on text tasks'; 262,144-token context, qwen3/qwen3_coder parsers, Think used the remaining context for output with no extra cap; sampling = Think recommendations (temperature 1.0, top_p 0.95, top_k 20, presence_penalty 1.5). Row 'IFBench · Strict prompt-level', printed 42.67. Strict (not loose) prompt-level accuracy. |
| LongBench v2 | 49.7% | maritaca-ai/S5-nano model card · maker-reported Maritaca own evaluation, pinned S5-nano card, 'Results by benchmark' table 'Think / Thinking', column 'S5-nano'. Card: 'Results for the original S5-nano checkpoint, evaluated in BF16 on text tasks'; 262,144-token context, qwen3/qwen3_coder parsers, Think used the remaining context for output with no extra cap; sampling = Think recommendations (temperature 1.0, top_p 0.95, top_k 20, presence_penalty 1.5). Row 'LongBench v2 · Accuracy', printed 49.70. Card limitations note: '503 questions ... confirmed context-overflow cases receive zero and remain in the evaluation population.' |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json