# Tiny Analysis > Independent comparison of open-weight language models under 100B parameters: a composite Small LM Index, benchmarks (with who measured each score), size, memory, API price, speed and licenses. 139 models, updated 2026-11-01. The Small LM Index (0-100) is the predicted average over six core benchmarks (MMLU-Pro, GPQA Diamond, Humanity's Last Exam, AIME, LiveCodeBench, SWE-bench Verified). Benchmarks a lab didn't report are estimated from its other scores, and a likely range is given. Independently measured scores replace self-reported ones where available. ## Best models by size ### Tiny (< 5B) - [Nanbeige4.2 3B](https://tinyanalysis.com/models/nanbeige4.2-3b): Nanbeige (Kanzhun), 4.2B, index 70.4, Apache-2.0 - [Spark-X2.6 4B](https://tinyanalysis.com/models/spark-x2.6-4b): SparkLLM (XHToken), 4.1B, index 67.5, Apache-2.0 - [Spark-X2.5 4B](https://tinyanalysis.com/models/spark-x2.5-4b): SparkLLM (XHToken), 4.1B, index 65.5, Apache-2.0 - [LFM3 1.7B](https://tinyanalysis.com/models/lfm3-1.7b): Liquid AI, 1.7B, index 60.0, LFM Open License - [Nemotron Labs Dory 1.0 4B](https://tinyanalysis.com/models/nemotron-labs-dory-4b): NVIDIA, 4B, index 60.0, OpenMDW-1.1 ### Small (5–15B) - [Qwen3.5 9B](https://tinyanalysis.com/models/qwen3.5-9b): Alibaba (Qwen), 9B, index 63.9, Apache-2.0, $0.028/1M tokens - [Granite 4.2 8B](https://tinyanalysis.com/models/granite-4.2-8b): IBM (Granite), 8B, index 63.5, Apache-2.0, $0.082/1M tokens - [Apriel 1.6 15B Thinker](https://tinyanalysis.com/models/apriel-1.6-15b-thinker): ServiceNow (Apriel), 14.9B, index 60.9, MIT - [Ministral 3 14B](https://tinyanalysis.com/models/ministral-3-14b): Mistral AI, 14B, index 60.0, Apache-2.0, $0.20/1M tokens - [Falcon H1R 7B](https://tinyanalysis.com/models/falcon-h1r-7b): TII (Falcon), 7.6B, index 57.5, Falcon LLM License ### Medium (15–40B) - [Nanbeige4.3 33B](https://tinyanalysis.com/models/nanbeige4.3-33b): Nanbeige (Kanzhun), 33.1B, index 77.8, Apache-2.0 - [Qwen3.8 27B](https://tinyanalysis.com/models/qwen3.8-27b): Alibaba (Qwen), 27B, index 77.0, Apache-2.0, $0.13/1M tokens - [Ternary Bonsai 2 27B](https://tinyanalysis.com/models/ternary-bonsai-2-27b): PrismML, 27.4B, index 74.6, Apache-2.0 - [Qwen3.6 27B](https://tinyanalysis.com/models/qwen3.6-27b): Alibaba (Qwen), 27B, index 74.0, Apache-2.0, $0.90/1M tokens - [Muse Glimmer 30B](https://tinyanalysis.com/models/muse-glimmer-30b): Meta, 29.6B, index 72.6, Apache-2.0, $0.53/1M tokens ### Large (40–100B) - [Nemotron 3.5 Bolt 74B-A9B](https://tinyanalysis.com/models/nemotron-3.5-bolt-74b-a9b): NVIDIA, 74B (9B active), index 73.3, OpenMDW-1.1 - [Kolibri-1](https://tinyanalysis.com/models/kolibri-1): Aleph Alpha, 78.1B (3.5B active), index 68.3, Apache-2.0 - [Qwen3-Next 80B-A3B Thinking](https://tinyanalysis.com/models/qwen3-next-80b-a3b-thinking): Alibaba (Qwen), 80B (3B active), index 63.3, Apache-2.0 - [K2-V2 Instruct](https://tinyanalysis.com/models/k2-v2): LLM360 / MBZUAI (K2), 70B, index 57.9, Apache-2.0 - [DeepSeek-R1-Distill-Llama-70B](https://tinyanalysis.com/models/deepseek-r1-distill-llama-70b): DeepSeek, 70.6B, index 50.1, MIT ## Best model for the job - [Best small open models for coding](https://tinyanalysis.com/best/coding): Qwen3.8 27B (83.5), Ternary Bonsai 2 27B (81.0), Qwen3.6 27B (80.6) - [Best small open models for math](https://tinyanalysis.com/best/math): Kolibri-1 (96.0), Ternary Bonsai 2 27B (95.8), Nanbeige4.3 33B (95.2) - [Best LLMs that run on an 8 GB GPU or laptop](https://tinyanalysis.com/best/8gb-gpu): Nanbeige4.2 3B (70.4), Spark-X2.6 4B (67.5), Spark-X2.5 4B (65.5) - [Best LLMs that run on a 16 GB GPU](https://tinyanalysis.com/best/16gb-gpu): Nanbeige4.2 3B (70.4), Spark-X2.6 4B (67.5), Spark-X2.5 4B (65.5) - [Best LLMs for a 24 GB GPU (RTX 4090, RTX 3090)](https://tinyanalysis.com/best/24gb-gpu): Nanbeige4.3 33B (77.8), Qwen3.8 27B (77.0), Ternary Bonsai 2 27B (74.6) - [Best tiny LLMs for phones and edge devices](https://tinyanalysis.com/best/phones-edge): Nanbeige4.2 3B (70.4), Spark-X2.6 4B (67.5), Spark-X2.5 4B (65.5) - [Best cheap LLM APIs (open models)](https://tinyanalysis.com/best/cheap-api): Qwen3.8 27B (77.0), Qwen3.6 35B-A3B (72.2), Gemma 4 31B (70.1) - [Best small open vision-language models](https://tinyanalysis.com/best/vision): Qwen3.8 27B (77.0), Ternary Bonsai 2 27B (74.6), Qwen3.6 27B (74.0) - [Best open models with an Apache 2.0 or MIT license](https://tinyanalysis.com/best/commercial-license): Nanbeige4.3 33B (77.8), Qwen3.8 27B (77.0), Ternary Bonsai 2 27B (74.6) - [Best European open language models](https://tinyanalysis.com/best/european): Kolibri-1 (68.3), Ministral 3 14B (60.0), Magistral Small 1.2 (55.9) - [Best American open models under 100B](https://tinyanalysis.com/best/american): Ternary Bonsai 2 27B (74.6), Nemotron 3.5 Bolt 74B-A9B (73.3), Muse Glimmer 30B (72.6) - [Best fully open models (open data and code)](https://tinyanalysis.com/best/fully-open): K2-V2 Instruct (57.9), OLMo 3.1 32B Think (57.0), OLMo 3 7B Think (43.4) ## Pages - [Leaderboard and chart explorer](https://tinyanalysis.com/): filter by region, size, license, reasoning; any two axes. - [Charts](https://tinyanalysis.com/charts) - [Compare models side by side](https://tinyanalysis.com/compare), with ready-made head-to-head pages such as /compare/-vs- - [Guides: best model for coding, math, 8/16/24 GB GPUs, phones, cheap APIs](https://tinyanalysis.com/best) - [Labs by region](https://tinyanalysis.com/vendors) - [Can I run it?](https://tinyanalysis.com/can-i-run): which models fit a given GPU, Mac, PC or phone, at what quantization, memory and speed. Each model page also lists Ollama, llama.cpp, LM Studio and MLX commands. - [About](https://tinyanalysis.com/about): independence and how every score is sourced and labelled. - [Methodology](https://tinyanalysis.com/methodology): how the index is computed and where numbers come from. - [All data as JSON](https://tinyanalysis.com/api/models.json)