Show HN: Mark 1x-9B – a 9B model that answers with interfaces, not paragraphs
21 days ago
- Saanora/mark-1x-9b is a 9B parameter model trained to emit Saanora's directive DSL for interactive UI (3D scenes, charts, quizzes, flashcards, etc.).
- It achieves 84% directive validity on first try (schema-free) and 87% restraint (prose when no interface helps), with multi-turn artifact edits at 12.3/14.
- Built on Qwen3.5-9B via LoRA r=32 merged to bf16, with 262,144 token context, 201 languages, and thinking mode inherited.
- Usage instructions are provided for Transformers, vLLM, SGLang, and Docker Model Runner, including code snippets and API calls.
- Self-reported benchmarks: IFEval 84.7, MMLU-Pro 80.6, GPQA Diamond 79.4, AIME 2025 78.3, AIME 2026 83.3, HMMT Feb 2026 63.6.
- Known limitations: weak exact-text constraints, tool use without harness reduces accuracy, need explicit system prompt, small label noise, and no vision support.
- Released under OpenMDW-1.0 license, derivative of Qwen3.5-9B (Apache-2.0), with required attribution.