OPEN//FRONTIER

REVIEWED · JUL 20, 2026

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.

14

models tracked

13

weights live

4

openness tiers

1M

largest context

MODEL LEDGER · 01

Frontier index

14 of 14 models

Provider benchmarks use different harnesses and are shown as evidence notes, never as a cross-model ranking.

Comparable properties of open-weight frontier models
CompareModelStatusScaleContextInputsReasoning + agentsOpennessSelf-hostingSource
Kimi K3Moonshot AI · Jul 16, 2026Weights Jul 272.8TUndisclosed active1MAdvertised
TextVision
Max at launch; lower efforts plannedCoding, tools, Agent SwarmLicense pendingTBD with weightsAPI today; weights promised Jul 27Official ↗
InklingThinking Machines · Jul 15, 2026Weights live975B41B active1M64K / 256K on Tinker
TextVisionAudio
Controllable effortCoding, tools, fine-tuning on TinkerUse-restrictedApache 2.0 + Model AUP≥600 GB quantized; ≥2 TB BF16Official ↗
GLM-5.2Z.ai · Jun 16, 2026Weights live744B40B active1MNative long-horizon target
Text
Off / high / maxTool calling, long-horizon engineeringPermissiveMITCluster class; BF16 and FP8Official ↗
Kimi K2.7 CodeMoonshot AI · Jun 12, 2026Weights live1T32B active256KNative
TextVision
Forced, preserved across turnsLong-horizon coding and toolsUse-restrictedModified MITLarge multi-GPU; native INT4Official ↗
Nemotron 3 UltraNVIDIA · Jun 4, 2026Weights live550B55B active1MRULER tested
Text
Off / regular / mediumStructured tools, delegation, recoveryPermissiveOpenMDW 1.18×H200/B200 or 16×H100 BF16Official ↗
Gemma 4 12BGoogle DeepMind · Jun 3, 2026Weights live12B12B dense active256KNative
TextVisionAudio
Configurable thinkingNative function calling, structured JSONPermissiveApache 2.0Laptop class; targets 16 GB memoryOfficial ↗
MiniMax M3MiniMax · Jun 1, 2026Weights live~428B~23B active1MNative
TextVisionVideo
Off / enabled / adaptiveCoding, computer use, long-horizon agentsUse-restrictedMiniMax CommunityCluster classOfficial ↗
Step 3.7 FlashStepFun · May 29, 2026Weights live196B + 1.8B vision~11B active256KNative
TextVision
Low / medium / highVisual search, tools, GUI workflowsPermissiveApache 2.0~120 GB minimum unified memoryOfficial ↗
Mistral Medium 3.5Mistral AI · Apr 28, 2026Weights live128B128B dense active256KNative
TextVision
Configurable effortFunction calling, JSON, codingUse-restrictedModified MITAs few as 4 GPUs (vendor)Official ↗
DeepSeek V4 ProDeepSeek · Apr 24, 2026Preview · weights live1.6T49B active1MThink Max recommends ≥384K
Text
Off / think / think maxCoding, reasoning, toolsPermissiveMITCluster class; mixed FP4 / FP8Official ↗
Qwen3.6 27BAlibaba Qwen · Apr 22, 2026Weights live27B27B dense active256KExtensible to ~1M
TextVisionVideo
Thinking / non-thinkingTools, repository reasoning, codingPermissiveApache 2.0Workstation/server; 8 GPUs for full 256KOfficial ↗
GLM-4.7-FlashZ.ai · Jan 19, 2026Weights live30B3B active~200K202,752 positions
Text
Reasoning with preserved thinkingTools and coding agentsPermissiveMITLocal/workstation friendlyOfficial ↗
gpt-oss-120bOpenAI · Aug 5, 2025Weights live117B5.1B active128KNative
Text
Low / medium / highTools, functions, structured outputPermissiveApache 2.0 + usage policySingle 80 GB GPUOfficial ↗
Llama 4 MaverickMeta · Apr 5, 2025Weights live400B17B active1MNative claim
TextVision
General reasoningTool support depends on runnerCustom licenseLlama 4 CommunityDGX-class FP8 deploymentOfficial ↗

COMPARE · 2/2

Inkling ↔ GLM-5.2

COMPARISON · 02

Inkling versus GLM-5.2

Property
Inkling
GLM-5.2
Status
Weights live
Weights live
Scale
975B total · 41B active
744B total · 40B active
Context
1M · 64K / 256K on Tinker
1M · Native long-horizon target
Modalities
Text + Vision + Audio
Text
Reasoning
Controllable effort
Off / high / max
Agents & tools
Coding, tools, fine-tuning on Tinker
Tool calling, long-horizon engineering
Openness
Use-restricted · Apache 2.0 + Model AUP
Permissive · MIT
Self-hosting
≥600 GB quantized; ≥2 TB BF16
Cluster class; BF16 and FP8
Best fit
Custom multimodal agents
Long-context coding agents
Benchmark note
Vendor: SWE-bench Verified 77.6
Vendor: Terminal-Bench 2.1 81.0

METHODOLOGY · 03

Open is not a binary.

01

Weights first

“Live” means a downloadable checkpoint exists. API access alone does not qualify.

02

License decoded

Permissive, use-restricted, custom, and pending terms are kept separate.

03

Claims stay sourced

Provider benchmarks remain labeled and are not normalized into a misleading leaderboard.

04

Hardware counts

A model is only “runnable” in context: laptop, workstation, or cluster-class.