LLM Benchmarks — Scores, Methodology & Live Rankings
What each benchmark measures, how models score, and the current rankings
IntPhys2
Video Understanding
IntPhys2 intuitive-physics video understanding benchmark, reported on the Muse Spark Open model card under Multimodal (xhigh reasoning effort).
VideoMMEv2
Video Understanding
VideoMME v2 multimodal video understanding benchmark, reported on the Muse Spark Open model card and in its HF evaluation-results metadata (dataset: MME-Benc...
FORTRESS adversarial (static) ARS
Safety
FORTRESS adversarial robustness benchmark, static attack setup, Attack Relevance/Refusal Score (ARS); lower is better. Reported on the Muse Spark Open model ...
FORTRESS adversarial (static) ORS
Safety
FORTRESS adversarial robustness benchmark, static attack setup, Offense Rate Score (ORS); lower is better. Reported on the Muse Spark Open model card Safety ...
StrongREJECT v2 (adaptive)
Safety
StrongREJECT v2 jailbreak benchmark with adaptive attacks; lower is better. Reported on the Muse Spark Open model card Safety section.
EnterpriseOps-Gym
General Agents
EnterpriseOps-Gym - agentic enterprise-operations benchmark, evaluated using its official repository per the card's footnote; models run at highest thinking ...
Terminal-Bench Lite
Agentic Coding
Lite variant of Terminal-Bench (terminal coding agent evaluations) as reported on the KRAFTON Raon-1 mini preview card; evaluated with judge gpt-5.6-luna. Di...
BrowseComp Plus
General Agents
BrowseComp Plus (long-document extension of BrowseComp) as reported on the KRAFTON Raon-1 mini preview card; search/agentic benchmark, judge gpt-5.6-luna. Di...
Chatbot Korean (internal)
Multilingual
KRAFTON internal benchmark measuring Korean colloquial ability in chatbot conversations across diverse situations; not publicly released. Reported on the KRA...
Artificial Analysis Coding Agent Index
Composite
Aggregator index (Artificial Analysis) for autonomous agentic coding performance; combines scores across coding-agent evals and reports tasks completed (%). ...
Artificial Analysis Intelligence Index v5
Composite
Aggregator index (Artificial Analysis) Intelligence Index version 5. Chart on the ML4 card: GLM-5.3 32, DeepSeek V4.1 Flash 1017 32, MiMo-V2.6-Pro 30, Muse S...
Terminal-Bench Science 0.1
Agentic Coding
Terminal-Bench Science task variant (leaderboard task id terminalbenchscience_0_1), evaluated by Artificial Analysis on the harborframework/terminal-bench-sc...
FrontierCode 2.0
Agentic Coding
Cognition private benchmark assessing end-to-end code quality and whether a repository maintainer would merge the agent's change. Mistral blog: Mistral Large...
Artificial Analysis Cyber Index
Composite
Aggregator index (Artificial Analysis) measuring a model's ability to discover vulnerabilities in a codebase, reproduce and validate them, and patch them wit...
FinWorkBench (Finch)
General Agents
Finance-workflow benchmark measuring creation and editing of spreadsheets for finance and accounting use cases. Blog: ML4 Finch 73.3%, second only to Kimi K3...
B3 Agent Security Benchmark (Attack Resistance)
General Agents
B3 Agent Security Benchmark: resistance of deployed agents to attacks such as system prompt exfiltration and malicious code injection. Blog: ML4 resists 97.1...
Malicious Cyber Request Refusal
Composite
Average refusal rate on malicious cyber-related prompts of JailbreakBench, StrongREJECT and AgentHarm (higher = safer). Blog: ML4 96.6%, the highest Mistral ...
SWE Refactor Bench
Agentic Coding
Agentic software refactoring benchmark; agent harness Claude Code 2.1.220 at 200K context. Introduced on the MiniMax M3.1-Flash release card (M3.1-Flash + fr...
RULER 128K
Long Context
RULER 128K - long context retrieval benchmark evaluated at 128K tokens. As evaluated on the LFM3-1.7B card: model served via vLLM in BF16 at its full 262,144...
LiveCodeBench Pro 25Q2
STEM & Reasoning
LiveCodeBench Pro 2025 Q2 aggregate over all difficulty splits. As evaluated on the LFM3-1.7B card: vLLM BF16 at full 262,144-token context, temperature 0.6,...
Daily-Omni
Multimodal
Daily-Omni as evaluated on the LFM3-Omni-2B card: Omni table, 'Video + audio understanding' group. Multiple-choice accuracy on video with its soundtrack. No ...
WorldSense
Multimodal
WorldSense as evaluated on the LFM3-Omni-2B card: Omni table, 'Video + audio understanding' group. Multiple-choice accuracy on video with its soundtrack. No ...
AV-SpeakerBench
Multimodal
AV-SpeakerBench as evaluated on the LFM3-Omni-2B card: Omni table, 'Video + audio understanding' group. Multiple-choice accuracy on video with its soundtrack...
OmniBench (overall)
Multimodal
OmniBench (overall) as evaluated on the LFM3-Omni-2B card: Omni table, 'Image + audio understanding' group. Image plus audio clip per question, scored with t...
OmniBench (speech)
Multimodal
OmniBench (speech) as evaluated on the LFM3-Omni-2B card: Omni table, 'Image + audio understanding' group. OmniBench subset whose audio is speech (same offic...
RealWorldQA (spoken)
Vision & Language
RealWorldQA (spoken) as evaluated on the LFM3-Omni-2B card: image benchmarks with each question given only as speech (Omni table, 'Spoken visual QA' group).
MMStar (spoken)
Vision & Language
MMStar (spoken) as evaluated on the LFM3-Omni-2B card: MMStar with each question given only as speech (Omni table, 'Spoken visual QA' group).
POPE
Vision & Language
POPE as evaluated on the LFM3-Omni-2B card: POPE F1 over adversarial, popular, and random subsets (image setting).
POPE (spoken)
Vision & Language
POPE (spoken) as evaluated on the LFM3-Omni-2B card: POPE F1 with each question given only as speech (Omni table, 'Spoken visual QA' group).
ERQA (spoken)
Vision & Language
ERQA (spoken) as evaluated on the LFM3-Omni-2B card: ERQA with each question given only as speech (Omni table, 'Spoken visual QA' group).
EmbSpatial (spoken)
Vision & Language
EmbSpatial (spoken) as evaluated on the LFM3-Omni-2B card: Embodied spatial QA with each question given only as speech (Omni table, 'Spoken visual QA' group).
SEEDBench (image)
Vision & Language
SEEDBench (image) as evaluated on the LFM3-Omni-2B card: image subset of SEED-Bench.
RoboSpatial
Spatial Intelligence
RoboSpatial as evaluated on the LFM3-Omni-2B card: RoboSpatial-Home with the official RoboSpatial-Eval rule.
English ASR WER
Multimodal
English ASR WER as evaluated on the LFM3-Omni-2B card: average word error rate (WER) in percent on the ESB / Open ASR benchmark suite; lower is better.
English TTS UTMOSv2
Multimodal
English TTS UTMOSv2 as evaluated on the LFM3-Omni-2B card: UTMOSv2 speech quality score on a 1-5 scale, higher is better.
VoiceBench
Multimodal
VoiceBench as evaluated on the LFM3-Omni-2B card: audio input, text output, 0-100 scale, higher is better.
Tool calling (internal eval)
Multimodal
Tool calling (internal eval) as evaluated on the LFM3-Omni-2B card: Liquid AI internal voice tool-calling evaluation, audio input, percentage, higher is better.
MMAR
Multimodal
MMAR as evaluated on the LFM3-Omni-2B card: MMAR audio understanding and reasoning, percentage, higher is better.
MME
Vision & Language
MME as evaluated on the LFM3-VL-2B card (vision benchmark, with-thinking setting): General category, with-thinking column. Card prints MME as a normalized pe...
SEED-Bench (image)
Vision & Language
SEED-Bench (image) as evaluated on the LFM3-VL-2B card (vision benchmark, with-thinking setting): General category, with-thinking column. Image subset of SEE...
MM-Vet (v1)
Vision & Language
MM-Vet (v1) as evaluated on the LFM3-VL-2B card (vision benchmark, with-thinking setting): General category, with-thinking column. MM-Vet v1 (original); not ...
MM-IFEval
Instruction Following
MM-IFEval as evaluated on the LFM3-VL-2B card (vision benchmark, with-thinking setting): General category, with-thinking column. Multimodal instruction follo...
MuirBench
Video Understanding
MuirBench as evaluated on the LFM3-VL-2B card (vision benchmark, with-thinking setting): General category, with-thinking column. Image/video difference under...
PerceptionBench
Vision & Language
PerceptionBench as evaluated on the LFM3-VL-2B card (vision benchmark, with-thinking setting): Perception category, with-thinking column. No inventory match.
HR-Bench (4K)
Vision & Language
HR-Bench (4K) as evaluated on the LFM3-VL-2B card (vision benchmark, with-thinking setting): Perception category, with-thinking column. High-resolution 4K su...
HR-Bench (8K)
Vision & Language
HR-Bench (8K) as evaluated on the LFM3-VL-2B card (vision benchmark, with-thinking setting): Perception category, with-thinking column. High-resolution 8K su...
ZoomBench (full + crop average)
Vision & Language
ZoomBench (full + crop average) as evaluated on the LFM3-VL-2B card (vision benchmark, with-thinking setting): Perception category, with-thinking column. Car...
CV-Bench
Vision & Language
CV-Bench as evaluated on the LFM3-VL-2B card (vision benchmark, with-thinking setting): Perception category, with-thinking column. Cambrian Vision-Centric (2...
CharXiv (descriptive, val)
Document Understanding
CharXiv (descriptive, val) as evaluated on the LFM3-VL-2B card (vision benchmark, with-thinking setting): Document/OCR & Chart category, with-thinking column...
LogicVista
STEM & Reasoning
LogicVista as evaluated on the LFM3-VL-2B card (vision benchmark, with-thinking setting): Math & logical reasoning category, with-thinking column. Multimodal...