Can you own it?

Intended & out-of-scope use

Deploy Soofi-S for industrial and regulated German and English work - technical and regulatory documents, code, and agentic workflows where European data-residency and auditability matter. It is the first model from the German Soofi consortium ("Sovereign Open Source Foundation Models"), coordinated by the KI Bundesverband and funded by the German Federal Ministry for Economic Affairs and Energy (BMWE) under the European IPCEI-CIS programme

• a ~30B hybrid Mamba-2/Transformer Mixture-of-Experts model (30B total, ~3B active per token) trained on ~27 trillion tokens of primarily German and English.

For adoption, deploy the Instruct-Preview (post-trained) variant; the base checkpoint is built for fine-tuning and is out of scope for customer-facing use. Because this is an early preview/beta, high-stakes, autonomous, or safety-critical decision-making is out of scope without the external control stack described in the Implement domain - treat it as a strong open foundation to pilot, not a hardened production endpoint.

Known limitations, bias & failure modes

• Early preview/beta. Treat capability, safety and tooling as still stabilising, not final.
• Incomplete safety hardening. The Instruct-Preview model card documents privacy and safety features as still incomplete, and there is no companion guard/classifier model.
• Custom architecture. The hybrid Mamba-2/MoE ships custom modelling code and must be loaded with trust_remote_code=True; runtime support is still maturing across serving stacks.
• Strong-open, not frontier-vs-closed. It is the strongest fully open model on English + German aggregates, but that is a claim relative to open baselines, not the largest closed labs.

The offsetting advantage is transparency: because the full per-source data accounting and code are published, you can inspect what went into the model rather than reasoning about an opaque artifact.

Openness tier & components

Soofi-S is aiming at the OLMo-style fully_open tier, but as of this grounding it has not landed there - so it sits at open_weights with Dimension 1 scored 3, not 5. What is genuinely delivered and read is the documentation: the pretraining report publishes full per-source data accounting (source identifiers, raw and effective token counts, epoch multipliers, and even sources evaluated but excluded) and commits the project to the Open Source AI Definition (OSAID 1.0). But that is a pledge, not a release: the base weights are in a closed-beta phase ("open model weights coming soon" on the soofi-project/Soofi-Pretraining README), the training and evaluation code is not yet released, and the open licence is unconfirmed (no SPDX; the LICENSE file returned HTTP 401 when read). A fully_open/5 rating requires all six components actually Open - downloadable weights under a permissive licence with runnable code - none of which can be verified against a read primary document today. The honest read is promised-and- well-documented, not delivered: revisit and re-score when the OSAID release lands. (A further caveat for later: ~1.3% of Phase 1 tokens, the commercially licensed Genios corpus, are reported in aggregate rather than being redistributable, so even the committed release clears OSAID 1.0 but is only nearly compliant with stricter "every token redistributable" open-data proposals.)

License terms & permitted use

Soofi-S is committed to a permissive, OSAID-compliant open license with unconditional commercial use and no field-of-use restriction - you may use, modify, redistribute and commercialise derivatives. Two honest caveats keep this from being a fully settled Apache-style story: the final license SPDX is not yet published, and the base weights are currently in a closed-beta access phase transitioning to the committed open release. This is why the legal dimension scores 4 rather than 5 - the openness intent and released artifacts are excellent, but the terms are still being finalised.

Supply-chain & provenance

Weights are distributed from the verified Soofi-Project org on Hugging Face in bf16 safetensors (no pickle requirement), accompanied by first-party GGUF and 3-bit quantizations, with per-file checksums. Because the data accounting, training code and intermediate checkpoints are all public, provenance is highly auditable - the checkpoint trust checklist scores 6/8. The two missing controls are cryptographic weight signing (Sigstore/model-signing) and SLSA build attestation, and the base weights sit behind a closed-beta access gate for now, which is why provenance is a 4, not a 5. No incidents or malicious-mirror findings are on record. Pin the exact revision and verify checksums on download.

EU AI Act posture

Soofi-S is a GPAI model but sits well under the 10²⁵-FLOPs systemic-risk threshold (~30B total / ~3B active, ~27T tokens), so no Article 55 regime applies. It is committed to a free/open-source release that would meet OSAID 1.0 and is not monetised, which would qualify it for the Article 53 open-source exemption from the Annex XI/XII technical-documentation duties - but the exemption cannot yet be relied on, because the open licence is not finalised (base weights in closed beta, no SPDX), so there is no confirmed free-and-open licence to point to today. The two obligations that survive - a copyright policy and a public training-content summary - are where Soofi-S is unusually strong: because it publishes full per-source data accounting, the training-content summary can be assembled directly from released artifacts (only the ~1.3% Genios slice is aggregate-only). Combined with an EU/Germany jurisdiction and training on sovereign German infrastructure, this makes Soofi-S one of the cleaner EU-AI-Act postures in the registry. A fine-tuner placing a derivative on the EU market may become a provider for that derivative but inherits an unusually complete upstream package.

Benchmarks & evaluation

OneHill did not run its own benchmarks this session; the picture below is aggregated from the Soofi pretraining technical report and independent coverage (ev-perf, ev-arxiv). The consistent finding: Soofi-S is the strongest fully-open model on aggregate English and German benchmarks, ahead of OLMo 3 32B, Apertus 70B, EuroLLM 22B and Alia 40B, matching dense 14-27B models and winning code aggregates in both languages among open base models - remarkable at only ~3B active parameters. This item is marked partial because exact per-task numbers are still being filled in on the public model cards (some are flagged "to be measured"); treat the leaderboard-style claims as third-party/in-class, not as a frontier-vs-closed claim.