[Model] Support for new “Sovereign AI Foundation Model" project models #16230
Description
Motivation
Requesting SGLang support for the new and future “Sovereign AI Foundation Model” project models so they can be served without vendor forks.
Summary of details below:
K-EXAONE-236B-A23B (LGAI-EXAONE/K-EXAONE-236B-A23B)
LG AI is maintaining an SGLang fork with support for this model.
- Type: LLM, MoE (sparse)
- Params: 236B total / 23B active per token; 128 experts, top-8 (+1 shared)
- Context: 256K (262,144) native long-context
- Attention: 3:1 hybrid (3× sliding-window + 1× global), window=128; NoPE (no RoPE)
- Modes: Think vs Non-Think
- Benchmarks: Published on model card (reasoning/agentic/long-context table)
VAETKI-112B-A10B (NC-AI-consortium-VAETKI/VAETKI)
Model doesn't seem to be fully public yet, as technical report and evals have yet to be released.
- Type: LLM, MoE
- Params: 112.2B total / 10.1B active per token; 128 experts, top-8
- Context: 128K
- Modes: Think vs Non-Think (tool-agent uses non-thinking; most other tasks use thinking)
- Benchmarks: TBA
A.X-K1 (skt/A.X-K1)
Public release for benchmarks of the model on Jan 4, 2026. States that it can be served with SGLang, will SGLang have day 1 support for A.X-K1?
- Type: LLM, MoE
- Params: 519B total / 33B active per token; 192 experts (+1 shared), top-8 (+1 shared)
- Context: 131,072
- Modes: Think vs Non-Think (single model)
- Benchmarks: Public release on Jan 4, 2026
Solar-Open-100B (upstage/Solar-Open-100B)
A vLLM fork is available.
- Type: LLM, MoE
- Params: 102.6B total / 12B active per token; 128 routed +1 shared, top-8
- Context: 128K
- Modes: Not specified
- Benchmarks: TBA
HyperCLOVAX-SEED-Think-32B (naver-hyperclovax/HyperCLOVAX-SEED-Think-32B)
Naver provides OmniServe, a production-ready multimodal inference system with OpenAI-compatible API.
- Type: Multimodal VLM (text/image/video in → text out); dense 32B
- Params: 32B (dense)
- Context: 128K
- Modes: Optional thinking mode
- Benchmarks: TBA