Is it good enough?

Capabilities & modalities

Soofi-S is a text-only language model built for German and English generation, code, and technical/agentic tasks in industrial and regulated settings. Per the pretraining report and independent coverage (ev-arxiv, ev-perf), it is the strongest fully-open model on aggregate English and German benchmarks - ahead of OLMo 3 32B, Apertus 70B, EuroLLM 22B and Alia 40B - and wins the code aggregates in both languages among open base models, all while activating only ~3B of 30B parameters per token. Choose Soofi-S when European sovereignty, auditability, German-language strength and long-context economics matter; it is a leading open model rather than a frontier-vs-closed contender.

Context window & long-context behaviour

Soofi-S supports contexts up to 256k tokens, and its hybrid Mamba-2 design keeps decode throughput flat from 4k to 256k (a near-constant inference cache) - a genuine differentiator for long-document and high-concurrency workloads. This item is marked partial because OneHill has not independently measured effective long-context recall this session; treat 256k as the architectural ceiling rather than a verified quality guarantee at full length.

Prompt format & chat template

The Instruct-Preview variant ships an embedded Jinja chat template (its default system prompt carries only a knowledge cutoff and date - no identity text). Apply it rather than hand-rolling role markers:

messages = [
    {"role": "system", "content": "Du bist ein hilfreicher Assistent."},
    {"role": "user", "content": "Fasse das folgende Dokument zusammen ..."},
]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

With GGUF, run llama-server --jinja so the embedded template (identity and tools) is applied verbatim. The base checkpoint has no chat template - it is a plain text-completion / fine-tuning model and should not be prompted as a chat assistant.

Language coverage

Soofi-S is a genuine bilingual German + English model: German was deliberately up-weighted in the ~27T-token training mix, and both languages are first-class supported targets (this is the model's core design goal, not incidental coverage). Other languages may appear but are not a supported target - evaluate per-language before relying on Soofi-S outside German/English.

Function / tool calling

Soofi-S has a native tool-call format embedded in the Instruct GGUF's Jinja template. To use it reliably, run under llama-server --jinja, which applies the model's own identity-and-tools template verbatim. Be aware that Ollama's template intentionally omits the tool portion, so tool calling there falls back to prompt scaffolding - this is why the item is marked partial: the capability exists and is documented, but its behaviour depends on which runtime template your stack applies.