Open-model programme · reviewed 29 July 2026

Use the strongest model that fits the boundary.

ANULUM evaluates serious open-source and open-weight candidates across compact nodes, professional systems and rack-scale infrastructure. Kimi K3 is the current flagship open-weight reference for maximum-capability evaluation—not the automatic answer to every workload.

Official sourcesSystem fitANULUM score pending

01Licence and weight custody02Workload and language fit03Hardware and operating fit04Measured acceptance

Open source and open weight are not synonyms.

Open source is reserved here for models whose exact artefacts and terms support that description. Open weight means weights can be obtained under stated terms, which may be custom, gated or use-restricted. Every engagement reviews the exact checkpoint, licence version, intended use and redistribution path.

Catalogue rule: inclusion means “worth evaluating”, not “endorsed”, “safe”, “compliant” or “production-ready”. Third-party quantisations enter only after provenance and integrity checks.

Current governed candidate families

The roster is broad enough to avoid vendor lock-in and narrow enough to remain source-reviewable.

Inspect the detailed catalogue →

Swiss option

Apertus

Multilingual, Swiss-developed candidates for controlled RAG, document and multimodal work where local provenance matters.

Use: confidential archives, multilingual retrieval, compact and large nodes.

Permissive candidates

gpt-oss · OLMo · DeepSeek V4

Reasoning, tool use and research-transparent or permissively licensed paths at very different system scales.

Use: agents, reasoning, coding, long context and controlled research.

Compact multimodal

Gemma · Qwen · Granite · Ministral

Smaller model families for document extraction, local assistants, vision and cost-sensitive private deployment.

Use: OCR, classification, office workflows, edge and single-node systems.

Large generalists

Mistral Small 4 · Llama 4 · GLM

Multimodal and agentic candidates between professional multi-GPU systems and rack-class deployments.

Use: multilingual assistants, visual documents, tools and higher-concurrency services.

Specialists

Devstral · Kimi Code · Granite Guardian

Purpose-shaped candidates belong in specialist lanes instead of silently becoming the default general model.

Use: software engineering, policy checks, guardrails and narrow automation.

Flagship reference

Kimi K3

Official weights and serving guidance make K3 a credible maximum-capability sovereign-evaluation candidate. It remains a rack programme with client-specific proof gates.

Open the K3 evaluation →

Benchmark the use case, not the brand.

Cited archive answers

Measure: retrieval recall, claim support, citation correctness, abstention and reviewer time on a frozen document set.

Professional drafting

Measure: factual defects, instruction adherence, source fidelity, redaction behaviour and human revision burden.

Coding and agents

Measure: held-out task completion, test pass rate, tool-schema validity, permission adherence, rollback and receipt completeness.

Multimodal documents

Measure: OCR and table accuracy, layout grounding, image-supported claims and failure visibility across real client formats.

Swiss multilingual work

Measure: German, English, French and Italian task quality, terminology stability, named-entity preservation and citation fidelity.

System performance

Measure: time to first token, sustained output, concurrency, memory, context cost, energy and recovery on the declared system.

Score status

Publisher results are evidence—not ANULUM results.

Official cards and reports establish publisher-reported benchmarks. ANULUM publishes a measured result only with the checkpoint hash, engine and settings, declared hardware, dataset or task version, run record and acceptance threshold.

Read the benchmark contract →

Primary family sources

Sources reviewed 29 July 2026. Model inventories and terms change; the exact checkpoint and licence are re-verified at engagement start.