# Under 9gb model comparison — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/43a50e9244b8e64d

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-16 17:39 ingested 2026-11-17 02:41

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xxvr5z/under_9gb_model_comparison/)

## retained text



```
Scores are lab-reported so take them with a grain of salt.

# Specs

|Model|Org|Params (total/active)|Context|Modality|License|
|:-|:-|:-|:-|:-|:-|
|K2-Horizon-7B|IFM|7B dense|512K|text|Apache-2.0|
|Granite 4.2-8B|IBM|8B dense (thinking)|128K (512K ext.)|text|Apache-2.0|
|Qwen3.5-9B|Qwen|9B hybrid (delta-net, thinking)|262K (\~1M ext.)|text + vision + video|Apache-2.0|
|MiMo-V2.6-Distill-Qwen-9B|Xiaomi|9B (SFT of Qwen3.5-9B)|\~262K (base)|text + vision (visual coding)|MIT|
|LFM2.5-8B-A1B|Liquid|8.3B / 1.5B active MoE|128K|text|LFM 1.0|
|Gemma 4 E4B|Google|4.5B effective / 8B total (PLE)|128K|text + image + audio|Apache-2.0|
|LFM3-VL-2B|Liquid|1.92B (1.73B backbone, thinking)|128K|text + image + video|LFM 1.0|

# Text benchmarks

Superscripts: a = K2 card (IFM harness), b = Liquid harness, c = MiMo card (Xiaomi harness); no mark = own card; e = from LFM3-1.7B, the text backbone; d/† = MiMo avg@3 / internal set. Bold = best in row (cross-harness caveat applies).

|Benchmark|K2-7B|Granite 8B|Qwen3.5-9B|MiMo-D-9B|LFM2.5-8B-A1B|Gemma 4 E4B|LFM3-VL-2B|
|:-|:-|:-|:-|:-|:-|:-|:-|
|Math||||||||
|AIME 2026|90.2||92.5||50.0|42.5||
|AIME 2025|90.3|86.7|56.1ᵇ||42.5|34.3ᵇ||
|HMMT Feb 2026|77.8|66.5ᵃ|65.7ᵃ|||||
|MATH-500|||||88.8|65.0ᵇ||
|Coding||||||||
|SWE-bench Verified|72.4|47.7|50.8ᵃ|61.1ᵈ||0.5ᵇ|43.4ᵉ|
|SWE-bench Pro||19.1||44.6ᵈ||||
|SWE-bench Multilingual||30.8|35.0ᵇ|||0.7ᵇ|23.0ᵉ|
|Terminal-Bench 2.1|44.6|20.6|29.2ᵃ|37.1||0.0ᵇ|12.9ᵉ|
|LiveCodeBench v6||73.2|65.6|||52.0|73.5ᵉ|
|SciCode|33.6|30.4ᵃ|27.5ᵃ|||25.0ᵇ|28.5ᵉ|
|Agentic||||||||
|AutomationBench|||15.1ᵇ|30.3||1.9ᵇ|41.3ᵉ|
|LiveClaw|||35.8ᵇ|||20.4ᵇ|45.6ᵉ|
|tau/ϴ-bench Banking|25.8|7.6ᵃ|7.0ᵃ|||4.1ᵇ||
|tau2-Bench|||79.1||88.1T / 39.8R|42.2||
|WebArena Verified|||28.9ᵇ|||17.1ᵇ|38.4ᵉ|
|BrowseComp / BrowseComp-Plus|59.0||38.0ᵇ|||21.0ᵇ|37.0ᵉ|
|PinchBench|||71.5ᵇ|||55.1ᵇ||
|Claw-Eval (EN avg)|||66.5ᵇ|||56.1ᵇ|68.6ᵉ|
|Toolathlon-Verified|||25.9ᶜ|35.2||||
|OfficeQA|||9.0ᶜ|19.5||||
|JobBench|||2.6ᶜ|18.3||||
|BFCL v4 (tool use)||52.4|66.1||49.7|48.3ᵇ|71.4ᵉ|
|ToolSandbox|||76.4ᵇ|||69.3ᵇ|80.4ᵉ|
|MCP-Atlas|||42.7ᵇ|||6.2ᵇ|32.6ᵉ|
|EnterpriseOps-Gym|||29.0ᵇ|||13.4ᵇ|28.2ᵉ|
|MiMo Code (mini) †|||19.5ᶜ|51.6||||
|MiMo Cyber (mini) †|||5.7ᶜ|31.3||||
|MiMo General (mini) †|||28.5ᶜ|62.2||||
|Knowledge||||||||
|GPQA Diamond|75.6|64.1|81.7|||58.6|63.4ᵉ|
|HLE|19.5|9.7ᵃ|14.9ᵃ|||4.8ᵇ|7.2ᵉ|
|MMLU-Pro||74.0|82.5|||69.4||
|AA-Omniscience Index|||\-28.0ᵇ||\-24.7|\-15.3ᵇ|0.5ᵉ|
|Instruction following||||||||
|IFEval|||91.5||91.8|90.0ᵇ|96.7ᵉ|
|IFBench|79.3|64.5|||56.5|44.6ᵇ|73.0ᵉ|
|Multi-IF|||||79.9|81.0ᵇ|89.1ᵉ|
|IFStruct|||89.2ᵇ|||78.5ᵇ|93.7ᵉ|
|Long context||||||||
|LCR / AA-LCR|68.0|43.3ᵃ|63.0|||39.3ᵇ|52.8ᵉ|
|RULER 128K||71.4|87.1ᵇ|||18.8ᵇ|65.6ᵉ|

# Vision benchmarks

|Benchmark|LFM3-VL-2B|Gemma 4 E4B|Qwen3.5-9B|MiMo-D-9B|
|:-|:-|:-|:-|:-|
|MMMU (val)|55.7||78.4||
|MMMU-Pro|45.7|52.6|70.1||
|MathVista (mini)|84.2||85.7||
|MATH-Vision|51.6|59.5|78.9||
|OCRBench v1|86.3||89.2||
|RefCOCO (avg)|90.5||89.7||
|ScreenSpot-Pro|55.5||65.2||
|RealWorldQA|79.3||80.3||
|MMStar|74.4||79.7||
|HallusionBench|65.6||69.3||
|MVBench (video)|69.4||74.4||
|Video-MME (subtitles)|69.0||84.5||
|MedXpertQA-MM||28.7|49.9||
|MiMo Visual Coding (mini) †|||61.7ᶜ|64.0|
```



document fetched 2026-11-17 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1xxvr5z/under_9gb_model_comparison
