One-line: The efficient member of the Jais-3 (Arabic-native K2 Horizon) fleet — 36B stored / 4B active MoE that posts the family’s best agentic numbers (SWE-bench Verified offline 72.8, Terminal-Bench 2.1 67.3, MCPMark 69.3) within a ~75 GB bf16 footprint (TP=2), so roughly Qwen3.8-27B-class repo work at MoE serving cost. All numbers vendor-reported; no third-party replication.
Active params: 4B active / 36B stored (recipe page: 5.95B / 37.44B) — MoE, so serving cost tracks active params while VRAM tracks stored params
Context length: 524,288 tokens (512K)
VRAM (fp16 rough): ~75 GB bf16 for all experts resident (IFM recipe: --tensor-parallel-size 2 with expert parallelism enabled); FP8 build ≈38 GB; GGUF and MLX builds published in the same org, so a single-GPU or Mac Studio path exists at reduced precision
Custom architecture:model_type: k2_horizon with remote code, and MoVA attention means vLLM/SGLang recipes (rather than stock transformers) are the practical path
Serving: vLLM ≥0.30 — --reasoning-parser k2_horizon, --tool-call-parser k2_horizon (json / xml / xml_typed, default xml), reasoning_effort per request (default high); language-dependent sampling (temp 1.0 English vs 0.7 Arabic in thinking mode); ≥32K output budget recommended
What makes it notable
The best agent-harness numbers IFM has published for an Arabic-native open model, at a footprint a two-GPU box can serve — and it beats its own 32B dense sibling on every agentic row except Apex-Agents. Read against Qwen3.8-27B it is a tool-selection-first model: MCPMark 69.3 vs 49.6 and BrowseComp-Plus 71.8 vs 53.0, while giving up 12.5 points on Terminal-Bench 2.1. Caveats: SWE-bench numbers are “offline” harness runs, all rows are vendor-reported with no independent replication, the benchmark suite is IFM’s own audited Arabic set, and the card’s param counts disagree with the recipe page’s (36B/4B vs 37.44B/5.95B).