Can you own it?
This page is a projection of the one entry record, the Use & modify and Transparency factors that Assess covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
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. The model is a ~30B hybrid Mamba-2/Transformer Mixture-of-Experts (30B total, ~3B active per token) trained on ~27 trillion tokens of primarily German and English.
Deploy the Instruct-Preview (post-trained) variant; the base checkpoint is built for fine-tuning and is out of scope for customer-facing use. This is an early preview/beta, so high-stakes, autonomous or safety-critical decision-making is out of scope without the external control stack described under Implement. Treat it as a strong foundation to pilot, not a hardened production endpoint.
Known limitations, bias & failure modes
Soofi-S is an early preview/beta, and capability, safety and tooling are still stabilising. The
Instruct-Preview model card documents privacy and safety features as still incomplete, and there
is no companion guard or classifier model. 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.
Third-party coverage calls Soofi-S the strongest "fully open" model on English and German aggregates, below the largest closed labs. That is a benchmark claim relative to open baselines, not a statement about access: the weights are gated and not an open release. Access needs publisher approval of an access request, and the closed beta is for selected partners. No release date or licence is published.
The offsetting advantage is documentation. The full per-source data accounting and code are published, so what went into the model can be inspected.
Openness tier & components
Soofi-S aims at the OLMo-style fully_open tier and is not there. It sits at gated_open, with
Dimension 1 scored 2. The weights are not openly available. Every repo in the Soofi-Project org
is manually gated. The base model card says "It is not an open release" and describes
"a closed beta with already selected partners". The licence tag reads closed-beta, and the preview
is "for research and development only". The project site says only that an open-source release
"is planned", with no date.
Marketing and binding text conflict. The paper title ("A Sovereign, Open-Source Foundation Model"), the organisation name and press coverage say open-source; the card, the tag and the gate say closed beta. The grade follows the binding text.
Documentation and code are delivered. The pretraining report publishes full per-source data
accounting (source identifiers, raw and effective token counts, epoch multipliers, and sources
evaluated but excluded). The Soofi-Pretraining repo holds the complete training codebase but
has no licence file, so it grants no reuse rights. The evaluation scores are a CC-BY-4.0 dataset.
The Open Source AI Definition (OSAID 1.0) commitment is a future-tense pledge ("will be
released") that has not been met since the July report. A fully_open/5 rating needs all six
components actually open; an anchor-3 rating needs open weights plus a licence, and neither
exists. Openness is promised and well documented, not delivered.
About 1.3% of Phase 1 tokens (the commercially licensed Genios corpus) are reported in aggregate and are not redistributable. The committed release would clear OSAID 1.0 but would only nearly satisfy stricter "every token redistributable" open-data proposals.
License terms & permitted use
No licence is in force. The licence tag reads closed-beta, no repo in the org (and not the
training-code repo) contains a LICENSE file, the card says the model "Will be released under a
permissive license", and the preview is limited to research and development. The consortium has
committed to a permissive, OSAID-compliant licence later, but no published text grants use,
modification, redistribution or commercial use. This is why the legal dimension scores 2:
documentation and jurisdiction are strong, and the missing licence is the binding gap. The schema
has no "unknown" commercial-use value, so the entry records allowed_with_conditions, which here
means only that no unconditional grant exists. Re-assess when a licence is published.
Supply-chain & provenance
Weights are distributed, to approved accounts only, from the verified Soofi-Project org on
Hugging Face in bf16 safetensors (no pickle requirement), with first-party GGUF, FP8 and 3-bit
quantizations and per-file checksums. The data accounting and training code are public, so
provenance is auditable. The intermediate Megatron checkpoint repo is gated like the rest.
The checkpoint trust checklist scores 6/8, a July score that is unverified. The two missing controls are cryptographic weight signing (Sigstore/model-signing) and SLSA build attestation. The base weights also sit behind a closed-beta access gate (every repo is manually gated), and the Hub reports no completed security scan for the base repo, so provenance scores 4, not 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 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. The exemption cannot yet be relied on, because no open licence is published (base weights manually gated in closed beta, licence tag closed-beta, no LICENSE file). There is no confirmed free-and-open licence to point to.
Two obligations survive the exemption: a copyright policy and a public training-content summary. Soofi-S is strong here. It publishes full per-source data accounting, so the training-content summary can be assembled directly from released artifacts (only the ~1.3% Genios slice is aggregate-only). With EU/Germany jurisdiction and training on sovereign German infrastructure, this is 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 and receives a complete upstream package.
Benchmarks & evaluation
Aggregated from the Soofi pretraining technical report and independent coverage (ev-perf,
ev-arxiv). The consistent finding is that 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. It
matches dense 14-27B models and wins code aggregates in both languages among open base models, at
only ~3B active parameters. Coverage is 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 and in-class, not as a frontier-vs-closed claim.
How this scores
The ownership factors this domain covers, drawn from the one entry record.
Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?
WeakA hard access gate with no licence in force. Every repo is manually gated, the card says it is not an open release and that the preview is for research and development only, and no LICENSE file exists. No published text grants use, fine-tuning or commercial use. The permissive release is a promise without a date.
TransparencyDo you know what it is: weights, training, behaviour, and legible terms?
ModerateThe report, the per-source data accounting and the training code are public and inspectable, which is more than most weight-only releases offer. The weights are not publicly inspectable (manual gate, closed beta) and the training code has no licence file, so transparency stays at moderate until the weights and licence are released.
Sources
The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.