🏆Tiny League 36 models · 95 benchmarks

The league table for small open models.

Benchmark scores for every recent small open-weight model worth comparing — read them side by side, see where each one is cheapest to run, and filter down to the ones that fit your hardware.

36models
95benchmarks
651data points
15with vision
Changelog
  • 6 Nov 2026Corrected Ornith-1.5-122B active parameters (10B active), fixing its decode-speed estimates.
  • 2 Nov 2026Added Nanbeige/Nanbeige4.3-33B; estimated decode speed, a Speed preset and sticky Models headers.
  • 28 Oct 2026v2 redesign: core heatmap, shareable filters, KV-cache column, HLE/IFBench scatter presets; added Spark-X2.6-1.7B, dropped KAT-Coder.
  • 27 Oct 2026Added LiquidAI/LFM3-1.7B; lowered the param floor to 1.5B.
  • 22 Oct 2026Added Sage-A1.5, GPQA/Terminal-Bench scatter charts, and τ²/τ³-bench scores.
  • 20 Oct 2026Added XHToken/Spark-X2.6-4B (4.1B dense) and dropped the superseded Spark-X2.5-4B. Added ornith-ai/Ornith-1.5-122B (125B MoE, 8B active, vision) and LiquidAI/LFM2.5-24B-A2B (24B MoE, 2B active). Removed the two oldest entries, Nemotron-3-Nano-Omni-30B-A3B-Reasoning and tencent/Hy-MT2-30B-A3B. The coverage heatmap now hides benchmark columns with fewer than two entries as soon as a filter is applied.
  • 15 Oct 2026Added nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 and nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16. Re-based the AA Intelligence Index on v4.3.2 and re-fetched every model.
  • 12 Oct 2026Added openbmb/MiniCPM-V-4.7-35B-A3B (35.2B MoE, vision) — first open-source on-device MoE VLM. Scores from the model card's eval figures (instruct, 4x visual-token compression).
  • 14 Sept 2026Added IFM/K2-Horizon-3.7B. All K2 Horizon models are now on Artificial Analysis: added the AA Intelligence Index (v4.3) to K2-Horizon-3.7B (16), K2-Horizon-7B (21) and K2-Horizon-MoVA-36B-A4B (26).
  • 13 Sept 2026Fixed duplicate benchmark scores (highest shown, conflicts logged), AA fetcher matching, and vision/param detection for new models.
  • 12 Sept 2026Added Agnes-AI/Agnes-3.0-Flash (33.1B dense, vision). Added a License column to the Models table.
  • 10 Sept 2026Added inclusionAI/Ling-3.0-flash-VL (124.8B MoE, vision) and meta-models/Muse-Glimmer-30B (29.8B dense, vision).
  • 9 Sept 2026Added nex-agi/Nex-N2.5-mini (35.1B MoE, vision). Added a model filter. Fixed the mobile layout (tables now scroll horizontally, heatmap keeps the model column pinned).
  • 8 Sept 2026Removed the Fara family (and JetBrains/Mellum2-12B); benchmark columns now show specific versions, grouped by family.
  • 7 Sept 2026Added openbmb/MiniCPM5-2B (2.5B dense), IFM K2-Horizon-MoVA-36B-A4B (36B MoE, 4B active) and K2-Horizon-7B (9B dense).
  • 4 Sept 2026Initial release with 26 open-weight models.

Models 36

Newest first. Click any header to sort; click again to flip. Active is the parameters a model uses per token (dense models use all of them); Context is its native maximum; Benchmarks is how many it reports. KV @128K is the cache needed to hold a 128K-token context (BF16, one sequence) — derived from the model's config.json and verified against the released weights, so it is a hardware cost, not a benchmark. Hover for the per-token rate. Decode fresh and Decode @128K are estimated single-stream output speeds on one NVIDIA RTX 4090 at half its 1008 GB/s bandwidth — before and after the KV read at 128K, memory-bandwidth bound. The sub-line is the share of fresh-context speed retained. Both are estimates, not measurements.

ModelParamsActiveContextKV @128KDecode freshDecode @128KTypeLicenseReleasedBenchmarks
LFM3-1.7B1.7B—262K1.50 GiB148 tok/s101 tok/s68% keptdenseother2026-10-2727
Nanbeige4.3-33B33.1B—524K26.0 GiB7.6 tok/s5.4 tok/s70% keptdenseApache 2.02026-10-2420
Spark-X2.6-4B4.1B—1M4.55 GiB62 tok/s39 tok/s63% keptdenseApache 2.02026-10-1927
Spark-X2.6-1.7B1.7B—1M1.77 GiB148 tok/s96 tok/s64% keptdenseApache 2.02026-10-1927
Ornith-1.5-122BVMoE125.1B10B262K3.00 GiB25 tok/s22 tok/s86% keptmoeMIT2026-10-1613
Sage-A1.5VMoE36B3B262K2.50 GiB84 tok/s58 tok/s69% keptmoeApache 2.02026-10-1522
NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16VMoE124.5B12B1M1.00 GiB21 tok/s20 tok/s96% keptmoeother2026-10-1236
LFM2.5-24B-A2BMoE24B2B131K2.50 GiB126 tok/s75 tok/s60% keptmoeother2026-10-1216
NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16VMoE74.9B9B262K1.00 GiB28 tok/s26 tok/s94% keptmoeother2026-10-0917
MiniCPM-V-4.7-35B-A3BVMoE35.2B3B262K2.50 GiB84 tok/s58 tok/s69% keptmoeApache 2.02026-10-0733
Agnes-3.0-FlashV33.1B—262K9.00 GiB7.6 tok/s6.6 tok/s87% keptdenseApache 2.02026-09-115
Nex-N2.5-miniVMoE35.1B—262K2.50 GiB7.2 tok/s6.9 tok/s96% keptmoeApache 2.02026-09-0816
MiniCPM5-2B2.5B—131K5.25 GiB101 tok/s47 tok/s47% keptdenseApache 2.02026-09-0633
K2-Horizon-MoVA-36B-A4BMoE37.4B4B524K24.0 GiB63 tok/s15 tok/s24% keptmoeApache 2.02026-09-019
K2-Horizon-7B9B—524K18.0 GiB28 tok/s14 tok/s48% keptdenseApache 2.02026-09-018
K2-Horizon-3.7B5.1B—524K18.0 GiB49 tok/s17 tok/s34% keptdenseApache 2.02026-09-018
Qwen3.8-Flash-NextVMoE180B6B262K3.00 GiB42 tok/s33 tok/s79% keptmoeother2026-08-2423
Ornith-1.5-9BV9.7B—262K4.00 GiB26 tok/s21 tok/s82% keptdenseMIT2026-08-1813
Ornith-1.5-35B-A3BVMoE36B3B262K2.50 GiB84 tok/s58 tok/s69% keptmoeMIT2026-08-1815
Ling-3.0-tinyMoE7.9B1.3B131K0.84 GiB194 tok/s144 tok/s74% keptmoeMIT2026-08-1017
Muse-Glimmer-30BV29.8B—131K1.70 GiB8.5 tok/s8.2 tok/s97% keptdenseApache 2.02026-08-0927
granite-4.2-30b29.3B—131K32.0 GiB8.6 tok/s5.4 tok/s63% keptdenseApache 2.02026-08-0719
granite-4.2-8b8.8B—131K20.0 GiB29 tok/s13 tok/s45% keptdenseApache 2.02026-08-0719
Qwen3.8-27BV27.8B—262K8.00 GiB9.1 tok/s7.9 tok/s87% keptdenseApache 2.02026-08-0526
Ling-3.0-flashMoE127.5B5.1B262K0.98 GiB49 tok/s45 tok/s91% keptmoeMIT2026-08-0225
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16MoE31.6B3B262K0.75 GiB84 tok/s74 tok/s88% keptmoeother2026-08-0114
LongCat-Flash-Lite-SparseMoE69.1B3B983K3.94 GiB84 tok/s49 tok/s59% keptmoeMIT2026-07-3129
Nanbeige4.2-3B4.2B—262K22.0 GiB60 tok/s16 tok/s26% keptdenseApache 2.02026-07-2121
Laguna-S-2.1MoE117.6B8B1M6.07 GiB32 tok/s22 tok/s71% keptmoeopenmdw-1.12026-07-136
DeepSeek-V4-Flash-DSparkMoE284B13B1M0.68 GiB19 tok/s19 tok/s97% keptmoeMIT2026-06-2722
Qwen-AgentWorld-35B-A3BVMoE34.7B3B262K2.50 GiB84 tok/s58 tok/s69% keptmoeApache 2.02026-06-221
Laguna-XS-2.1MoE33.4B3B262K5.06 GiB84 tok/s44 tok/s53% keptmoeopenmdw-1.12026-06-204
diffusiongemma-26B-A4B-itVMoE25.8B4B262K2.69 GiB63 tok/s47 tok/s75% keptmoeApache 2.02026-06-0914
North-Mini-Code-1.0MoE30.5B3B500K3.53 GiB84 tok/s53 tok/s63% keptmoeApache 2.02026-06-0511
LFM2.5-8B-A1BMoE8.5B1.5B128K1.47 GiB168 tok/s110 tok/s66% keptmoeother2026-05-2812
gemma-4-12B-itV12B—262K2.31 GiB21 tok/s19 tok/s92% keptdenseApache 2.02026-05-2316

Scores heatmap 95 benchmarks

How to read this

Each cell is one model×benchmark score. Colour is scaled within a column (pale → deep = lower → higher), so each benchmark is read on its own scale; the medal cell is the best model on that benchmark. A column is only shown when at least two models report that exact benchmark — versions are never pooled. The Core view keeps the benchmarks with 5+ models; All shows every column. Click a header to rank models by it.

lowerhigherbest on benchmarkVvisionMoEMixture-of-Experts
Modelτ³-bench21τ²-bench10Terminal-Bench2.120Terminal-Bench2.05Terminal-Bench43GPQA Diamond26HLE25SWE-bench Pro22SciCode21IFBench20SWE-bench Verifi…20AA-LCR17SWE-bench Multil…17AIME202611AIME20254BrowseComp15LiveCodeBenchv612BFCLv412HMMTFeb 202610HMMTFeb 20252MCP-Atlas12MMLU-Pro9ClawEval8AA-Omniscience A…7DeepSWE4DeepSWE1.13IMO AnswerBench7MRCRv24MRCR3GDPval-AA-V26IFEval6MathVision6Multi-IF6NL2Repo-Bench6AA-Omniscience I…5AA-Omniscience N…5CharXiv5CritPt4EnterpriseOps-Gy…4GDPval4JobBench4LongBench-v24MATH5004MMMU Pro4OSWorld-Verified4OmniDocBench 1.54PinchBench4SWE Atlas4Toolathlon4Toolathlon-Verif…4WebArena-Verifie…4WideSearch4Agents' Last Exa…3AutomationBench3Codeforces ELO3GDPval-AA3GraphWalks3GraphWalks Exten…3RULER3SkillsBench3Vision2Web3AndroidWorld2Apex Shortlist2Arena-Hard-V22ArtifactsBench2BabyVision2BigBench Extra H…2BirdBench2BrowseComp-zh2CMMLU2Claw-Gym2ClawEval-MM2CoWorkBench2DeepResearch Ben…2ERQA2GAIA2GDP.pdf2Gaokao 20262LIFEbench2LiveClaw2LiveCodeBench Pro25Q22MMLU-ProX lite2MMMLU2MedXPertQA MM2OCRBench 22PAgentBench2ProfBench2RealWorldQA2RecreationBench2SWE-MM2SWebBench2SuperGPQA2ToolSandbox2VitaBench2VitaBench2.02
Qwen3.8-27BV48.0·73··89.230.861.7·79.5······90.3········42.2·····94.6·42.3··90.2···33.4···84.391.1····64.8·42.9·······62.981.9···85.6·····57.470.7·65.5············85.947.138.6·····
Qwen3.8-Flash-NextVMoE45.4····91.735.962.5·81.3··81···91.9········58.7·····95.7·48.1··90.6···55.7·······73.5···51.2·······6484.5·········64.473.9·72.3············88.549.9······
Qwen-AgentWorld-35B-A3BVMoE·······························································································
gemma-4-12B-itV·69···78.85.2······77.5··72····77.2·····43.4···79.7···········69.1·0.2········1659···········53···············83.448.7···········
diffusiongemma-26B-A4B-itVMoE·56.2···73.211.9······69.1··69.1····77.6·····32···70.5···········54.3·0.3········1429···········47.6···············81.549···········
DeepSeek-V4-Flash-DSparkMoE39.4··56.9·88.145.152.6··79·73.3··73.2··94.8·6986.2····88.4·78.7···················47.8·····30521395······85.7································
Laguna-S-2.1MoE··70.2····59.4····78.5···········40.4······················46.249.7··············································
Laguna-XS-2.1MoE···37.5···47.6··70.9·63.1··················································································
LFM2.5-8B-A1BMoE·88.1·······56.5···5042.5··49.7·····8.7······91.8·79.9·-24.763.5······88.8····················································
Nanbeige4.2-3B···44.1·87.417.846.935.654.663.658.7····72.5·82.8·57.8·52.2···67.3············74.3······························65··33.4·····················
North-Mini-Code-1.0MoE6.437.4·36·75.711.140.238.257.567.6·····70.3··············································································
granite-4.2-30b62·29.2··66.4·33.338.877.257·41.9·89.2·75.861.4·89.2·77.6·················1225··················90.0····67.9···41.9·············66.6····42.9········
granite-4.2-8b58.1·20.6··64.1·19.136.179.347.7·30.8·86.7·73.252.4·78.3·74.0·················1189··················81.0····65.2···41.1·············61.1····41.2········
LongCat-Flash-Lite-SparseMoE·96.0·33.7·69.5·40.6··68.24859.365.7·48.6··41.5·45.679.2····49.4·44.7············53.696.8··············66.3··········61.984.5·······················21.7·
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16MoE9.3·24.6··75.411.7·32.671.951.65239.3··37.0·····81.9·······832················85.4················································
Ornith-1.5-9BV··47··86.430.547.5··70.6·54.4··56.4····54.2·66.5··········32.4·············20.6·41.2·59.5···········································
Ornith-1.5-35B-A3BVMoE··68.5··89.233.459.6··79·71.4··67.6····70.2·72.5·22········46.2·············39.848.7··67.8···········································
Ling-3.0-tinyMoE20.8·27.7··73.49.3·24.263.6·58.7·····62.770.3····8.5··71.0·····83.2··69.5···················772········47.9·············62.3················
Ling-3.0-flashMoE28·57···22.756.6·74.5·65.172.493.2·82·7387·65.5·····83.7·90.8···87.7··················73.6···1107···44.8····77·············77.3················
K2-Horizon-MoVA-36B-A4BMoE26.8·58.6··80.825.2·38.9··66.3···········18.8···········69.2·2.1·························································
K2-Horizon-7B25.8·39.1···18.6·31.6·70.668···59··73.3············································································
MiniCPM5-2B20.897.18.6··70.28.914.426.366.346.459·86.586.539.769.166.663.8··70.8·······19.686.7·71.8········43.794.6·························43.5·59.2····88.7····68··········40.8···
Nex-N2.5-miniVMoE··73.4····43.8·······83.4·········36.1···1446··········28.5···71.2····54.663.4·········52.9····························25.5·····
Muse-Glimmer-30BV23.5·51.7··83.52251.243.6777680·94.7······75.5········953······78.8······7465.975.8·············44.3···································
Agnes-3.0-FlashV·····85.0··38.174.2·68.3···········23·······································································
K2-Horizon-3.7B17.7·25.1··65.412.9·25.9·68.6······50.970.5············································································
MiniCPM-V-4.7-35B-A3BVMoE·84.7·······47.7···76.5··56.660.1···83.5·········74.2····75.6······69.7·························90.2··············68.5······61.2···
NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16VMoE18.973.266.7·189.531.851.847.170.971.168.667.7··7888.2······26.224.6····1211·78.2··4.0·80.810.9······51.1···········89.7·····88.4··29.6··········6.6·······68.1··········
NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16VMoE17.0·53.9·088.828.7·44.5··67.4···70····55.6··21.8·····979.0····-19.256.1·4.3········90.3·····························2.8··················
Spark-X2.6-4B34.786.9···77.117.954.336.875.862.867.364.791.2·53.1·71.881.3·62.4·69.9···75.265.6··93.6·······33.6··53.3··············92.234.2···················134.9·······74.1····71.7···30.8
Ornith-1.5-122BVMoE··81.2···4876.953.8·84.5·76.4··78.5····74.6·79.4··········51.2···············70.8·75.2···········································
LFM2.5-24B-A2BMoE8.0····74.47.4·32.774.751······65.0············96.8·84.4·-24.9·······96.0···83.2···26.8··27.7·························48.1············80.7··
Sage-A1.5VMoE······55.161.244.381.6·8172.7··82.8·····················20········90·····19.1····················47.5·95.1·················45.4·
LFM3-1.7B68.0·12.9··63.47.2·28.573.043.452.823···73.571.4··32.6·68.67.1······96.7·89.1·0.592.9··28.2···········38.4··41.3····65.6····················45.631.7···········80.4··
Spark-X2.6-1.7B21.385.3···56.18.726.720.165.559.84857.370.1·41.7·55.248.8·41.8·66.9···47.156.1··89.4·······26.1··40.4··············85.125.7···················118.9·······58.8····53.9···22.6
Nanbeige4.3-33B48.8·80.9·23.889.5·60.352.6··84·95.2········78.3·65.5········53····4184.844.8······50.8·70.4···54.2·····61.1···································

Scatter one chart at a time

Pick a preset, or Custom to plot any two metrics against each other. The chart follows the filters, so it always shows the models currently in view. Hover a point for its values; click to open the model card.

Comparisons 48 benchmark charts

A leaderboard per benchmark: the top 10 models reporting it are drawn when at least four do, within the current filters. Bars are sorted and labelled, and the leader is filled in solid.