Weights first
“Live” means a downloadable checkpoint exists. API access alone does not qualify.
OPEN-WEIGHT INTELLIGENCE · EVIDENCE FIRST
The frontier you can actually run.
Open-weight models, compared without the hand-waving. See what is downloadable, what the license really permits, and what it takes to self-host.
34
models tracked
31
weights live
20
U.S. weights live
1M
largest context
Origin follows the primary pretraining organization, not contributor citizenship, investor location, or hosting region. Downloadable weights do not imply a permissive license. Provider benchmarks use different harnesses and are not a cross-model ranking.
| Compare | Model | Status | Scale | Context | Inputs | Reasoning + agents | License terms | Openness level | Self-hosting | Source |
|---|---|---|---|---|---|---|---|---|---|---|
| Muse Glimmer 30BMeta Superintelligence Labs · United States · Aug 10, 2026 | Weights live | 30B30B dense active | Not statedNo context-window size in the release announcement | Long-horizon agent execution with retry trainingLocal agents, function calling, local coding, and LLM-as-a-judge | PermissiveApache 2.0 | L1 · Weights | Under 20 GB quantized; designed for 24 GB or 32 GB memory | Official ↗ | ||
| Qwen3.8-MaxAlibaba Qwen · China · Announced · Aug 2, 2026 | Announced · weights next week | 2.4TUndisclosed active | TBDNot stated in supplied announcement | System-level autonomous planningAutonomous coding and professional workflows | License pendingTo be published with weights | Not yet open | Hosted model; downloadable weights pending$2/M input · $6/M output · $0.25/M implicit cache · pricing source ↗ | Official ↗ | ||
| BTL-3Bad Theory Labs · United States · Jul 23, 2026 | LoRA adapter live | 27B base requiredRank-32 LoRA adapter active | 256K32K launch benchmark context | Thinking-mode codingRepository work and structured tool use | PermissiveApache 2.0 adapter | L1 · Weights | Requires pinned Qwen3.6-27B base checkpoint | Official ↗ | ||
| Nanbeige 4.2 3BNanbeige · China · Jul 22, 2026 | Weights live | 4B incl. embeddings3B non-embedding active | 256K262,144 positions | Thinking model with agent post-trainingTool use, office work, and code agents | PermissiveApache 2.0 | L1 · Weights | 22 shared layers evaluated in two passes | Official ↗ | ||
| Solar Open 2Upstage · South Korea · Jul 22, 2026 | Weights live | 250B15B active | 1MNative hybrid attention | Long-horizon reasoningOffice, document, and coding agents | Custom licenseUpstage Solar License | L1 · Weights | Vendor minimum: 4× H200 | Official ↗ | ||
| Laguna S 2.1Poolside · United States · Jul 21, 2026 | Weights live | 118B8B active | 1MUp to 1M | Thinking / non-thinkingAgentic coding and long-horizon work | PermissiveOpenMDW-1.1 | L1 · Weights | Quantized builds available; full model remains memory-heavy | Official ↗ | ||
| Motif-3-BetaMotif Technologies · South Korea · Jul 21, 2026 | Beta · weights live | ~314B~13B active | 256KNative | Reasoning with custom architectureTool calling and general agents | Use-restrictedNon-commercial research license | L1 · Weights | Custom code; tested by vendor on H200 and B200 | Official ↗ | ||
| Antares-1BCisco Foundation AI · United States · Jul 21, 2026 | Weights live | 1B1B dense active | 128KGranite 4.0 backbone | SFT then GRPO for terminal explorationVulnerability localization in repositories | PermissiveApache 2.0 | L1 · Weights | Small checkpoint; terminal agent loop required | Official ↗ | ||
| Kimi K3Moonshot AI · China · Jul 16, 2026 | Weights Jul 27 | 2.8TUndisclosed active | 1MAdvertised | Max at launch; lower efforts plannedCoding, tools, Agent Swarm | License pendingTBD with weights | Not yet open | API today; weights promised Jul 27 | Official ↗ | ||
| InklingThinking Machines · United States · Jul 15, 2026 | Weights live | 975B41B active | 1M64K / 256K on Tinker | Controllable effortCoding, tools, fine-tuning on Tinker | Use-restrictedApache 2.0 + Model AUP | L1 · Weights | ≥600 GB quantized; ≥2 TB BF16 | Official ↗ | ||
| GLM-5.2Z.ai · China · Jun 16, 2026 | Weights live | 744B40B active | 1MNative long-horizon target | Off / high / maxTool calling, long-horizon engineering | PermissiveMIT | L1 · Weights | Cluster class; BF16 and FP8 | Official ↗ | ||
| Kimi K2.7 CodeMoonshot AI · China · Jun 12, 2026 | Weights live | 1T32B active | 256KNative | Forced, preserved across turnsLong-horizon coding and tools | Use-restrictedModified MIT | L1 · Weights | Large multi-GPU; native INT4 | Official ↗ | ||
| Nemotron 3 UltraNVIDIA · United States · Jun 4, 2026 | Weights live | 550B55B active | 1MRULER tested | Off / regular / mediumStructured tools, delegation, recovery | PermissiveOpenMDW 1.1 | L1 · Weights | 8×H200/B200 or 16×H100 BF16 | Official ↗ | ||
| Gemma 4 12BGoogle DeepMind · United States · Jun 3, 2026 | Weights live | 12B12B dense active | 256KNative | Configurable thinkingNative function calling, structured JSON | PermissiveApache 2.0 | L1 · Weights | Laptop class; targets 16 GB memory | Official ↗ | ||
| MiniMax M3MiniMax · China · Jun 1, 2026 | Weights live | ~428B~23B active | 1MNative | Off / enabled / adaptiveCoding, computer use, long-horizon agents | Use-restrictedMiniMax Community | L1 · Weights | Cluster class | Official ↗ | ||
| Step 3.7 FlashStepFun · China · May 29, 2026 | Weights live | 196B + 1.8B vision~11B active | 256KNative | Low / medium / highVisual search, tools, GUI workflows | PermissiveApache 2.0 | L1 · Weights | ~120 GB minimum unified memory | Official ↗ | ||
| Mistral Medium 3.5Mistral AI · France · Apr 28, 2026 | Weights live | 128B128B dense active | 256KNative | Configurable effortFunction calling, JSON, coding | Use-restrictedModified MIT | L1 · Weights | As few as 4 GPUs (vendor) | Official ↗ | ||
| DeepSeek V4 ProDeepSeek · China · Apr 24, 2026 | Preview · weights live | 1.6T49B active | 1MThink Max recommends ≥384K | Off / think / think maxCoding, reasoning, tools | PermissiveMIT | L1 · Weights | Cluster class; mixed FP4 / FP8 | Official ↗ | ||
| Qwen3.6 27BAlibaba Qwen · China · Apr 22, 2026 | Weights live | 27B27B dense active | 256KExtensible to ~1M | Thinking / non-thinkingTools, repository reasoning, coding | PermissiveApache 2.0 | L1 · Weights | Workstation/server; 8 GPUs for full 256K | Official ↗ | ||
| Trinity-Large-ThinkingArcee AI · Datology · Prime Intellect · United States · Apr 1, 2026 | Weights live | 398B~13B active | 512KExtended from 8K pretraining | Native thinking tracesTool calling and multi-step agents | PermissiveOpenMDW-1.1 | L1 · Weights | Cluster-class sparse MoE | Official ↗ | ||
| GLM-4.7-FlashZ.ai · China · Jan 19, 2026 | Weights live | 30B3B active | ~200K202,752 positions | Reasoning with preserved thinkingTools and coding agents | PermissiveMIT | L1 · Weights | Local/workstation friendly | Official ↗ | ||
| Olmo 3 32B ThinkAi2 · United States · Nov 19, 2025 | Weights live | 32B32B dense active | 65KYaRN extension from 8K | Long chain-of-thoughtMath, code, and research workflows | PermissiveApache 2.0 | L3 · Weights + algorithms + training data | Downloadable checkpoint; runner-dependent | Official ↗ | ||
| Granite 4.0 H 1BIBM · United States · Oct 28, 2025 | Weights live | 1.5B1.5B hybrid active | 128KMamba2 + attention | General instruction followingCompact enterprise workflows | PermissiveApache 2.0 | L1 · Weights | Small downloadable checkpoint | Official ↗ | ||
| Apriel H1 15B ThinkerServiceNow · United States · Oct 28, 2025 | Weights live | 15B15B dense active | 65KTarget; runtime-dependent | Thinker SFTEnterprise reasoning workflows | PermissiveMIT | L1 · Weights | Downloadable checkpoint; runner-dependent | Official ↗ | ||
| Marin 32B BaseStanford · Marin Community · United States · Oct 25, 2025 | Weights live | 32B32B dense active | 4KBase checkpoint | Base model; no chat tuningResearch and downstream post-training | PermissiveApache 2.0 | L1 · Weights | Downloadable base checkpoint | Official ↗ | ||
| LFM2-VL-3BLiquid AI · United States · Oct 22, 2025 | Weights live | 3B2.6B LM + 0.4B vision active | 32KText context | Lightweight multimodal reasoningInstruction following and agentic flows | Use-restrictedLFM Open License v1.0 | L1 · Weights | Small multimodal checkpoint | Official ↗ | ||
| Moondream 3 PreviewMoondream · United States · Sep 11, 2025 | Preview · weights live | 9B~2B active | 32KVision-language preview | Visual question answeringCaptioning, pointing, and vision tasks | Use-restrictedBSL 1.1 + Additional Use Grant | L1 · Weights | Downloadable preview; third-party service limits | Official ↗ | ||
| VaultGemma 1BGoogle · United States · Sep 5, 2025 | Weights live · gated | 1B1B dense active | 1KPretrained base model | Base model; no chat tuningResearch and downstream tuning | Custom licenseGemma Terms | L1 · Weights | Small checkpoint; terms acceptance required | Official ↗ | ||
| Nemotron Nano 9B v2NVIDIA · United States · Aug 18, 2025 | Weights live | 9B9B hybrid active | 128KMamba + Transformer | Reasoning on / offEfficient local reasoning | Custom licenseNVIDIA Open Model License | L1 · Weights | Small downloadable checkpoint | Official ↗ | ||
| gpt-oss-120bOpenAI · United States · Aug 5, 2025 | Weights live | 117B5.1B active | 128KNative | Low / medium / highTools, functions, structured output | PermissiveApache 2.0 + usage policy | L1 · Weights | Single 80 GB GPU | Official ↗ | ||
| SmolLM3 3BHugging Face · United States · Jul 8, 2025 | Weights live | 3B3B dense active | 128K64K trained; YaRN to 128K | Thinking / non-thinkingLightweight multilingual agents | PermissiveApache 2.0 | L2 · Weights + algorithms | Small downloadable checkpoint | Official ↗ | ||
| Phi-4 Mini Flash ReasoningMicrosoft · United States · Jun 19, 2025 | Weights live | 3.8B3.8B hybrid active | 64KHybrid SambaY architecture | Math-focused reasoningStructured logic and code workflows | PermissiveMIT | L1 · Weights | Small downloadable checkpoint | Official ↗ | ||
| Llama 4 MaverickMeta · United States · Apr 5, 2025 | Weights live | 400B17B active | 1MNative claim | General reasoningTool support depends on runner | Custom licenseLlama 4 Community | L1 · Weights | DGX-class FP8 deployment | Official ↗ | ||
| Reflection model (pending)Reflection AI · United States · No model released | Weights pending | TBDTBD | TBDNo public checkpoint | TBDOpen-model commitment only | License pendingTBD with release | Not yet open | No downloadable artifact | Official ↗ |
COMPARE · 0/2
Select two models
METHODOLOGY
Open is not a binary.
Three levels of open
Level 1 has open weights. Level 2 also has open algorithms. Level 3 also has open training data. The index assigns the highest evidenced level.
Claims stay sourced
Provider benchmarks remain labeled and are not normalized into a misleading leaderboard.
Hardware counts
A model is only “runnable” in context: laptop, workstation, or cluster-class.