Sub-9GB open model bake-off: Qwen3.5-9B most balanced, IFM K2-7B tops math and coding
parepeg · reddit · 2026-11-17
A Reddit user compiled a benchmark comparison of seven sub-9GB locally runnable models (mostly lab-reported scores): IFM K2-Horizon-7B, IBM Granite 4.2-8B, Qwen3.5-9B (hybrid delta-net, text+vision+video), Xiaomi MiMo-V2.6-Distill-9B, Liquid LFM2.5-8B-A1B (MoE), Google Gemma 4 E4B, and Liquid LFM3-VL-2B.
Key takeaways:
- Math: Qwen3.5-9B leads AIME 2026 at 92.5, K2-7B close at 90.2
- Coding: K2-7B dominates SWE-bench Verified at 72.4; Gemma 4 E4B scores 0.5, and 0.0 on Terminal-Bench
- Agentic/tool use: 2B LFM3-VL beats several larger models on LiveClaw and ToolSandbox; Qwen3.5-9B hits 79.1 on tau2-Bench
- Knowledge: Qwen3.5-9B tops GPQA Diamond (81.7) and MMLU-Pro (82.5)
- Vision: Qwen3.5-9B leads MMMU-Pro at 70.1
Overall Qwen3.5-9B is the most well-rounded; K2-7B excels at math/coding; tiny models show strong agentic value per parameter.
More from Models
- Open-source models now serve 30% of tokens, growing as fast as closed frontier AI — ivan_bezdomny · 2026-11-17
- OpenAI's flurry of blog posts sparks speculation of a model release today — imjustnewatai · 2026-11-17
- Polymarket claims GPT-6 Astra helped physicist crack 50-year particle physics problem — Polymarket · 2026-11-17
- Sam Altman says live: "We didn't have AGI, now we do" — imjustnewatai · 2026-11-17
- User says OpenAI refused $570 refund for auto-renewal with no explanation — OverLordGoldDragon · 2026-11-17
- Gemini spots the trick in a handwritten integral troll that stumps ChatGPT — vedantmisra · 2026-11-17