Apertus v1.5 8B
Deployable
Swiss, multilingual and compact, with text, image and audio input. A candidate for controlled RAG after exact task evaluation.
Model intelligence
Choose the model after the workload.
A model is deployable only when its weights, licence, serving engine, memory envelope and behaviour all fit. This catalogue separates those facts — and dates every conclusion.
Official model factsCalculated memory plansRegistry verified 26 July 2026 · next review by 26 August 2026
Explore the deployment catalogue
Filter by the system you want to own, not by a leaderboard headline.
12 models shown
Apertus v1.5 70B
Deployable
A Swiss large-model candidate with text, image and audio input; exact multimodal and long-context workloads require measurement.
gpt-oss-20b
Deployable
A compact reasoning and tool-use candidate. Active parameters affect compute, not the full weight-residency requirement.
gpt-oss-120b
Deployable
The official single-80-GB fit is a weight/runtime claim, not a multi-user throughput promise. Concurrency is benchmarked separately.
Qwen3.5 27B
Deployable
A compact multimodal option for document and vision workflows, subject to exact quantisation and context evaluation.
Kimi Linear 48B-A3B
Deployable
Interesting for long-context private archives. Published cache and latency advantages remain vendor/research evidence until measured on the delivery stack.
Kimi K2
Large-system
Rack-class despite 32B active parameters: the expert weights must remain resident. The official deployment guide is the sizing anchor.
Kimi K2.5
Large-system
Multimodal and more memory-efficient than FP8, but still a multi-GPU system whose context and concurrency must be proven.
Kimi K2.6
Large-system
A current multimodal checkpoint with image input. Local video is not claimed: the official card describes video support as experimental and API-only.
Kimi K2.7 Code
Coding specialist
A coding-agent derivative with preserved reasoning traces and image input. It is evaluated as a specialist, not assumed to be the default archive model.
DeepSeek V3
Large-system
A large MoE candidate where the custom model licence and validated multi-GPU serving path fit the client's governance boundary.
Kimi K3
Readiness watch
We do not advertise K3 as an on-prem model. Eligibility begins only after official weights, licence, hashes, engine support and a reproducible benchmark exist.
Kimi without the fog
The Kimi name spans very different deployment realities. Linear can fit a professional node when quantised. K2 through K2.7 are large-system programmes with distinct general, multimodal and coding roles. K3 remains a watch item, not an authorised local deployment claim.
Parameter honesty: active parameters estimate compute per token. Total resident weights, KV cache, runtime buffers, context and concurrency determine memory.
Primary evidence
- Official Moonshot AI model catalogue ↗
- Official Kimi K2 deployment guidance ↗
- Official Kimi K2.6 model card ↗
- Official Kimi K2.7 Code model card ↗
- Swiss AI Apertus v1.5 70B model card ↗
- Official gpt-oss release and deployment notes ↗
A source link does not constitute an ANULUM quality endorsement. Model selection follows evaluation on the client's tasks and constraints.