llmrun
Knowledge

LiveBench 2026 Research Leaderboard

LiveBench 2026 Research measures finding and combining information to answer research questions. It belongs to the 2026 LiveBench release (first snapshot September 2, 2026), a redesigned LiveBench in which each of seven categories is scored by a single task and questions are refreshed to limit training-set contamination. It is a different test from the original LiveBench boards, so the scores are not comparable.

Source: livebench15 open models ranked+33 proprietaryData through Sep 2026

Open modelsAll models

Open models ranked on LiveBench 2026 Research

# shows rank among open models / rank overall (including proprietary).

#ModelScoreVRAM @ Q4
1 / 6DeepSeek V4.1 Flash · 763.2B
85.3
~458 GB
2 / 16Muse Spark Open 816B A42B · 822.0B
80.3
~543 GB
3 / 18Kimi K3 · 2779.9B
79.0
~1674 GB
4 / 19MiMo V2.6 Pro RL · 1020B
76.7
~613 GB
5 / 22Smaug Flash · 304.2B
74.1
~609 GB
6 / 24GLM 5.3 · 753.3B
73.2
~456 GB
7 / 27NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 · 124.5B
70.9
~82 GB
8 / 28GLM 5.3 Flash · 321.3B
70.1
~195 GB
9 / 31DeepSeek V4 Flash Vision Exp · 304.6B
69.2
~183 GB
10 / 33MiniMax M3 · 427.0B
68.0
~257 GB
11 / 39Inkling · 952.4B
59.9
~572 GB
12 / 40Qwen3.8 27B · 27.8B
59.4
~17 GB
13 / 43Gemma 4 31B IT · 31.3B
44.4
~20 GB
14 / 45NVIDIA Nemotron 3 Ultra 550B A55B BF16 · 560.5B
43.2
~370 GB
15 / 47NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 · 31.6B
40.0
~19 GB

Score vs model size

Which models give the most quality for their size — the ones worth running locally.

100B1Tmodel size (log scale) →85.340.0Muse Spark Open 816B A42B · 822B · 80.3Kimi K3 · 2.8T · 79.0MiMo V2.6 Pro RL · 1T · 76.7GLM 5.3 · 753B · 73.2GLM 5.3 Flash · 321B · 70.1DeepSeek V4 Flash Vision Exp · 305B · 69.2MiniMax M3 · 427B · 68.0Inkling · 952B · 59.9Gemma 4 31B IT · 31B · 44.4NVIDIA Nemotron 3 Ultra 550B A55B BF16 · 561B · 43.2NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 · 32B · 40.0Qwen3.8 27B · 28B · 59.4Qwen3.8 27BNVIDIA Nemotron 3.5 Super VL 120B A12B BF16 · 124B · 70.9NVIDIA Nemotron 3.5 S…Smaug Flash · 304B · 74.1Smaug FlashDeepSeek V4.1 Flash · 763B · 85.3DeepSeek V4.1 Flash
Each dot is a model. Up = higher score, left = smaller (easier to run locally). The dashed line marks the efficiency frontier — the best score you can get at each size or smaller.
  • Qwen3.8 27B, 28B, score 59.4 — on the efficiency frontier (best score at its size or smaller).
  • NVIDIA Nemotron 3.5 Super VL 120B A12B BF16, 124B, score 70.9 — on the efficiency frontier (best score at its size or smaller).
  • Smaug Flash, 304B, score 74.1 — on the efficiency frontier (best score at its size or smaller).
  • DeepSeek V4.1 Flash, 763B, score 85.3 — on the efficiency frontier (best score at its size or smaller).

LiveBench 2026 Research: frequently asked questions

What is the best open LLM on LiveBench 2026 Research?
DeepSeek V4.1 Flash is the top open model on LiveBench 2026 Research, scoring 85.3. Among all models tested — including proprietary ones — it ranks #6. The top model overall is Claude Opus 5 (Anthropic) at 88.9.
What's the best LiveBench 2026 Research model you can run on a 24 GB GPU?
Qwen3.8 27B is the highest-scoring open model that fits in 24 GB at 4-bit quantization (~17 GB), scoring 59.4 on LiveBench 2026 Research.
Can open models match proprietary models on LiveBench 2026 Research?
Not quite on LiveBench 2026 Research: the strongest proprietary model (Claude Opus 5) scores 88.9, ahead of the best open model (DeepSeek V4.1 Flash) at 85.3 — but you can run the open one yourself.

Scores aggregated from livebench. llmrun does not run this benchmark — see the source for methodology, or the about benchmarks for what it measures.