Skip to content
AI Atlas
Open model frontier

Open models

Every canonical model whose weights can be downloaded, described by measurable properties — licence permissions, parameters, context, release, benchmark ranks, providers, estimated local fit. Openness is a set of observed dimensions here, never a score.

1,062 downloadable models

Open sourceWeights and code published under an OSI-approved licence that allows commercial use and derivatives.

10

Open weightsWeights downloadable under a permissive or Creative Commons licence allowing commercial use; code or data may be missing.

970

Restricted weightsWeights downloadable, but the licence restricts commercial use, hosting, derivatives or field of use (community, research and RAIL licences).

0

Downloadable totalUniverse = canonical models whose weights can be downloaded (open-weights, open-source, restricted-weights). hardware_fit values are ESTIMATES (see /methodology); best_results are ranks inside each benchmark's primary comparability group — no composite score.

1,062

New in 30 daysDownloadable models with a release date inside the last 30 days (release_date claim; models without one are not counted).

264

Openness explorer

Categories and licence permissions

Left: the downloadable universe by openness category. Right: what the licences on this page allow, counted per category — allowed · restricted · unknown.

Top licences: Apache License 2.0 (358) · Other (unclassified licence) (157) · MIT License (105) · Creative Commons Attribution-NonCommercial 4.0 (22) · Apple Sample Code / ML Research License (15) · Llama 3.2 Community License (12)

Category · this pageCommercial useRedistributionDerivativesHosting
Open source n=1100100100100
Open weights n=4914035140351403513036
restricted-weights n=0————

Counts read: allowed · restricted · unknown / unclassified — over the 50 models on this page, from each licence's stated terms (ontology).

Downloadable models

1,062 models · sorted by best benchmark rank

Permissions, in order:commercial use· redistribution· derivatives· hosting allowed restricted unknown

ModelLicence · permissionsParamsContextReleaseBest resultsProvidersFit 64 GB @4bit · 128 GB @8bit Estimated
Z.ai GLM 5.2Open weightsZ.ai (Zhipu AI) · textMIT753.3B1M16 Jun 2026tau2-bench#1terminal-bench#22livebench-2#382from $2 out✗ too large433.7 GB4bit✗ too large866.8 GB8bit
MiniMax M2.5Open weightsMiniMax · textOther228.7B204.8K12 Feb 2026swe-bench-verified#2swe-bench-multilingual#6tau2-bench#241from $1.2 out✗ too large132.0 GB4bit✗ too large263.5 GB8bit
GLM 4.7 FlashOpen weightsZ.ai (Zhipu AI) · textMIT31.2B200K19 Jan 2026tau2-bench#3ifbench#125terminal-bench#1631from $0.40 out✓ fits18.4 GB4bit✓ fits36.4 GB8bit
MiniMax-M3Open weightsMiniMax · image, text, videoOther427B1.05M2 Jun 2026ifbench#3gpqa#27terminal-bench#463from $1.2 out✗ too large246.1 GB4bit✗ too large491.6 GB8bit
GLM 5Open weightsZ.ai (Zhipu AI) · textMIT753.9B204.8K11 Feb 2026swe-bench-multilingual#4swe-bench-verified#6tau2-bench#81from $1.92 out✗ too large434.0 GB4bit✗ too large867.4 GB8bit
Nemotron 3 UltraOpen weightsNVIDIA · text—unclassified—262.1K4 Jun 2026ifbench#4livebench-2#51terminal-bench#721from $3.13 out——
Step 3.7 FlashOpen weightsStepFun · image, text, video—unclassified—262.1K28 May 2026tau2-bench#4terminal-bench#75mmmu-pro#771from $1.15 out——
DeepSeek-V4.1-FlashOpen weightsDeepSeek · image, text—unclassified—1M10 Sept 2026livebench-2#5artificial-analysis-intelligence-index#39scicode#562from $0.60 out——
Kimi K3Open weightsMoonshot AI · image, text, videoOther2.78T1.05M13 Jun 2026scicode#6livebench-2#10gpqa#143from $11.55 out✗ too large1599.0 GB4bit✗ too large3197.4 GB8bit
GLM 5.3Open weightsZ.ai (Zhipu AI) · textOther753.3B1.31M25 Aug 2026scicode#7artificial-analysis-intelligence-index#20livebench-2#253from $2.2 out✗ too large433.7 GB4bit✗ too large866.8 GB8bit
Kimi K2.5Open weightsMoonshot AI · image, textOther1.03T262.1K1 Jan 2026swe-bench-multilingual#7swe-bench-verified#10tau2-bench#161from $2.25 out✗ too large591.0 GB4bit✗ too large1181.4 GB8bit
GLM 5.1Open weightsZ.ai (Zhipu AI) · textMIT753.9B204.8K3 Apr 2026tau2-bench#9ifbench#23terminal-bench#411from $3.04 out✗ too large434.0 GB4bit✗ too large867.4 GB8bit
nemotron-cascade-2-30b-a3bOpen weightsNVIDIA—unclassified—1M—ifbench#9terminal-bench#166tau2-bench#2000——
MiMo-V2.5-ProOpen weightsXiaomi · text—unclassified—1.05M22 Apr 2026ifbench#10tau2-bench#34terminal-bench#411from $0.87 out——
DeepSeek V3.2Open weightsDeepSeek · textMIT685.4B163.8K1 Dec 2025swe-bench-verified#11swe-bench-multilingual#12tau2-bench#641from $0.40 out✗ too large394.6 GB4bit✗ too large788.7 GB8bit
deepseek-v4-flash-0420Open weightsDeepSeek—unclassified—1M—ifbench#12tau2-bench#27terminal-bench#750——
glm-5-non-reasoningOpen weightsZ.ai (Zhipu AI)—unclassified—200K—tau2-bench#12terminal-bench#58ifbench#1500——
glm-5-1-non-reasoningOpen weightsZ.ai (Zhipu AI)—unclassified—200K—tau2-bench#13terminal-bench#75humanitys-last-exam#1290——
Qwen3.5 397B A17BOpen weightsQwen · image, text, video—unclassified—262.1K16 Feb 2026ifbench#14tau2-bench#20terminal-bench#532from $3.5 out——
Qwen3.8 2.4T A95BOpen weightsQwen · text—unclassified—1.05M12 Aug 2026gpqa#14artificial-analysis-intelligence-index#33humanitys-last-exam#392from $6 out——
deepseek-v4-pro-0424Open weightsDeepSeek—unclassified—1M—tau2-bench#15ifbench#21terminal-bench#300——
GLM 4.7Open weightsZ.ai (Zhipu AI) · textMIT358.3B204.8K22 Dec 2025tau2-bench#16ifbench#84terminal-bench#1051from $1.75 out✗ too large206.6 GB4bit✗ too large412.6 GB8bit
Kimi K2.6Open weightsMoonshot AI · image, textOther1.03T262.1K14 Apr 2026tau2-bench#16ifbench#24terminal-bench#372from $4 out✗ too large591.0 GB4bit✗ too large1181.4 GB8bit
Kimi K2 ThinkingOpen weightsMoonshot AI · textOther1.03T262.1K4 Nov 2025swe-bench-verified#19tau2-bench#47ifbench#831from $2.5 out✗ too large590.7 GB4bit✗ too large1180.9 GB8bit
MiniMax M2Open weightsMiniMax · textOther228.7B204.8K22 Oct 2025swe-bench-verified#20ifbench#50tau2-bench#861from $1.02 out✗ too large132.0 GB4bit✗ too large263.5 GB8bit
deepseek-v4-flash-0420-highOpen weightsDeepSeek—unclassified—1M—tau2-bench#20ifbench#41terminal-bench#610——
Qwen3.6 35B A3BOpen weightsQwen · image, text, videoApache-2.036B3B active262.1K15 Apr 2026tau2-bench#24mmmu-pro#78terminal-bench#831from $0.90 out✓ fits21.2 GB4bit✓ fits41.8 GB8bit
Qwen3.8 FlashOpen weightsQwen · image, text, video—unclassified—1M26 Aug 2026livebench-2#24artificial-analysis-intelligence-index#35gpqa#372from $0.47 out——
GLM 4.6Open weightsZ.ai (Zhipu AI) · text—unclassified—204.8K30 Sept 2025swe-bench-verified#25terminal-bench#124tau2-bench#1401from $1.75 out——
qwen3-coder-480b-a35b-instructOpen weightsAlibaba Group · text—unclassified480B35B active256K22 Jul 2025swe-bench-verified#25terminal-bench#176tau2-bench#2260✗ too large276.5 GB4bit✗ too large552.5 GB8bit
Kimi K2 0711Open weightsMoonshot AI · textOther1.03T131.1K11 Jul 2025aider-polyglot#26swe-bench-verified#32tau2-bench#1871from $2.3 out✗ too large590.7 GB4bit✗ too large1180.9 GB8bit
GLM 4.5Open weightsZ.ai (Zhipu AI) · textMIT358.3B131.1K20 Jul 2025swe-bench-verified#27terminal-bench#163gpqa#2231from $2.2 out✗ too large206.6 GB4bit✗ too large412.6 GB8bit
R1Open weightsDeepSeek · textMIT684.5B64K20 Jan 2025aider-polyglot#271from $2.5 out✗ too large394.1 GB4bit✗ too large787.7 GB8bit
mimo-v2-flash-reasoningOpen weightsXiaomi—unclassified—256K—tau2-bench#27ifbench#112terminal-bench#1310——
Devstral 2Open weightsMistral AI · text—unclassified—262.1K9 Dec 2025swe-bench-verified#28scicode#148terminal-bench#1761from $2 out——
MiniMax M2.7Open weightsMiniMax · textOther228.7B204.8K9 Apr 2026ifbench#28terminal-bench#58scicode#712from $1.2 out✗ too large132.0 GB4bit✗ too large263.5 GB8bit
Qwen3.5-122B-A10BOpen weightsQwen · image, text, video—unclassified—262.1K25 Feb 2026ifbench#28tau2-bench#42mmmu-pro#801from $2.08 out——
DeepSeek V3 0324Open sourceDeepSeek · textMIT684.5B163.8K24 Mar 2025aider-polyglot#29scicode#126terminal-bench#2041from $1 out✗ too large394.1 GB4bit✗ too large787.7 GB8bit
GLM 5.3 FlashOpen weightsZ.ai (Zhipu AI) · image, text, videoMIT321.3B1.31M25 Aug 2026artificial-analysis-intelligence-index#29livebench-2#43gpqa#503from $0.25 out✗ too large185.3 GB4bit✗ too large370.0 GB8bit
Gemma 4 31BOpen weightsGoogle · image, text, videoApache-2.031.3B262.1K11 Mar 2026ifbench#30terminal-bench#72scicode#952from $0.34 out✓ fits18.5 GB4bit✓ fits36.5 GB8bit
Qwen3.5-27BOpen weightsQwen · image, text, video—unclassified—262.1K25 Feb 2026ifbench#30tau2-bench#39mmmu-pro#781from $1.56 out——
qwen3-8-27b-non-reasoningOpen weightsAlibaba Group—unclassified—256K—livebench-2#30mmmu-pro#125scicode#1410——
Step 3.5 FlashOpen weightsStepFun · text—unclassified—262.1K29 Jan 2026tau2-bench#31ifbench#95terminal-bench#1001from $0.30 out——
deepseek-v4-flash-0420-non-reasoningOpen weightsDeepSeek—unclassified—1M—tau2-bench#31terminal-bench#91ifbench#1990——
DeepSeek V4 Flash (0731)Open weightsDeepSeek · textMIT304.2B1.31M31 Jul 2026livebench-2#343from $0.08 out✗ too large175.4 GB4bit✗ too large350.3 GB8bit
Mistral Medium 3.5Open weightsMistral AI · image, textMIT-Modified—256K30 Apr 2026tau2-bench#34ifbench#79terminal-bench#951from $3.75 out——
Qwen3.6 27BOpen weightsQwen · image, text, videoApache-2.027.8B262.1K21 Apr 2026tau2-bench#34livebench-2#55terminal-bench#832from $2 out✓ fits16.5 GB4bit✓ fits32.5 GB8bit
deepseek-v4-pro-0424-highOpen weightsDeepSeek—unclassified—1M—tau2-bench#34terminal-bench#50ifbench#580——
glm-4-7-non-reasoningOpen weightsZ.ai (Zhipu AI)—unclassified—200K—tau2-bench#34terminal-bench#119ifbench#1540——
Kimi K2.7 CodeOpen weightsMoonshot AI · image, textOther1.03T262.1K11 Jun 2026terminal-bench#36livebench-2#49tau2-bench#682from $3.5 out✗ too large591.0 GB4bit✗ too large1181.4 GB8bit

Best results = rank inside each benchmark's primary comparability group; several dimensions are shown side by side and never combined. Fit columns are estimates (64 GB device at 4-bit, 128 GB at 8-bit, 8K context) — method on /methodology. unclassified = the raw licence label is not yet mapped in the ontology.

MethodUniverse = canonical models whose weights can be downloaded (open-weights, open-source, restricted-weights). hardware_fit values are ESTIMATES (see /methodology); best_results are ranks inside each benchmark's primary comparability group — no composite score. /methodology