Runtime as featured inForbesRead the article

The best open source AI models, ranked and compared

The top open-weight models for AI agents, with specs, licenses, and where each one is from.

Short answer: as of November 11, 2026, DeepSeek V4.1 Flash is the best open model for most agents. Kimi K3.1 Pro is the most capable when cost matters less. To self-host next to sensitive data, start with Qwen3.8 27B.

Updated · 18 models 12 from China 5 from the US 1 from France

En español:MéxicoColombiaChileEm português:Brasil

The best open source LLMs for agents

Ranked for agent work: tool use, context, license, and cost at volume.

2Moonshot AI logo

Kimi K3.1 Pro

Moonshot AI China

Most capable Kimi K3.1 Pro License

Moonshot's post-trained Kimi K3 posts the top open-weight Terminal-Bench 4.0 score in its report, within a point of Claude Opus 5.5.

Context
1M tokens
Params
2.8T · 104B active
Input
text, image
7Meta logo

Muse Spark Open

Meta US

Best US model Muse Model License

Meta's open-weight Muse Spark, built for coding and agent work with a 1M-token context.

Context
1M tokens
Params
816B · 42B active
Input
text, image, audio, video
8Mistral AI logo

Mistral Large 4

Mistral AI France

Best European model Mistral Ecosystem Open-Weight License

Mistral's largest model, a 1T-parameter multimodal MoE that ties GLM-5.3 on Artificial Analysis's coding agent index with fewer tokens per task.

Context
1M tokens
Params
1T · 52B active
Input
text, image
10Moonshot AI logo

Kimi K2.7 Code

Moonshot AI China

Best for coding Modified MIT

A coding-focused Kimi tuned for long-horizon software engineering.

Context
256K tokens
Params
1T · 32B active
Input
text, image
11NVIDIA logo
Best for documents and screens OpenMDW-1.1

NVIDIA's vision-language agent model, with 12B active parameters and a 1M-token context.

Context
1M tokens
Params
120B · 12B active
Input
text, image, video
13MiniMax logo

MiniMax M3

MiniMax China

Best for long multimodal inputs MiniMax Community License

A natively multimodal model with sparse attention built for million-token contexts.

Context
1M tokens
Params
428B · 23B active
Input
text, image, video
15Meta logo

Muse Glimmer

Meta US

Best for local agents Apache 2.0

A 30B agent model distilled from Muse Spark that runs on one 24GB to 32GB GPU or a Mac.

Context
128K tokens
Params
30B
Input
text, image
16NVIDIA logo

Nemotron 3.5 Bolt

NVIDIA US

Best on a 24GB GPU OpenMDW-1.1

A 74B model with 9B active that NVIDIA also ships in a 2-bit build sized for 24GB.

Context
256K tokens
Params
74B · 9B active
Input
text, image, video
17Qwen (Alibaba) logo

Qwen3.8 Flash Next

Qwen (Alibaba) China

Most efficient long context Qwen Community License

An experimental preview of the architecture that will underpin Qwen4.

Context
256K tokens
Params
180B
Input
text, image

Benchmark scores are reported by each model provider on its Hugging Face model card. They are not independently verified by Runtime.

Open source models compared

Filter by license, input, or origin. Sort by context or release date.

Filter:Origin:18 of 18 models
OriginParametersLicenseInputGood for
1DeepSeek logoDeepSeek V4.1 FlashBest overallChina552B / 8B1M tokensMITtext, imageOct 2026Transaction monitoringKYC and KYB checksPayments customer support
2Moonshot AI logoKimi K3.1 ProMost capableChina2.8T / 104B1M tokensKimi K3.1 Pro Licensetext, imageNov 2026Chargebacks and disputes
3Z.ai logoGLM-5.3Best for long agent runsChina753B1M tokensGLM-5.3 LicensetextAug 2026Payment reconciliationChargebacks and disputesMerchant underwriting
4Moonshot AI logoKimi K3Best for video inputChina2.8T1M tokensKimi K3 Licensetext, image, videoJun 2026Partner-bank RFIsMerchant underwriting
5MiniMax logoMiniMax M3.1 FlashBest for finance agentsChina428B / 23B1M tokensMiniMax Community Licensetext, image, videoOct 2026Partner-bank RFIsChargebacks and disputes
6DeepSeek logoDeepSeek V4 ProBest for reasoningChina1.6T1M tokensMITtextAug 2026Payment reconciliationMerchant underwritingAML and compliance investigations
7Meta logoMuse Spark OpenBest US modelUS816B / 42B1M tokensMuse Model Licensetext, image, audio, videoOct 2026
8Mistral AI logoMistral Large 4Best European modelFrance1T / 52B1M tokensMistral Ecosystem Open-Weight Licensetext, imageSep 2026
9Z.ai logoGLM-5.3 FlashBest valueChina320B / 18B1M tokensMITtext, imageAug 2026Partner-bank RFIsTransaction monitoringKYC and KYB checks
10Moonshot AI logoKimi K2.7 CodeBest for codingChina1T / 32B256K tokensModified MITtext, imageJun 2026
11NVIDIA logoNemotron 3.5 Super VLBest for documents and screensUS120B / 12B1M tokensOpenMDW-1.1text, image, videoOct 2026
12Qwen (Alibaba) logoQwen3.8 2.4T-A95BBest for multilingualChina2.4T / 95B256K tokensQwen3.8-Max LicensetextAug 2026AML and compliance investigationsPayments customer support
13MiniMax logoMiniMax M3Best for long multimodal inputsChina428B / 23B1M tokensMiniMax Community Licensetext, image, videoJun 2026
14Qwen (Alibaba) logoQwen3.8 27BBest to self-hostChina27B256K tokensApache 2.0text, image, videoAug 2026Payment reconciliationKYC and KYB checks
15Meta logoMuse GlimmerBest for local agentsUS30B128K tokensApache 2.0text, imageAug 2026
16NVIDIA logoNemotron 3.5 BoltBest on a 24GB GPUUS74B / 9B256K tokensOpenMDW-1.1text, image, videoOct 2026
17Qwen (Alibaba) logoQwen3.8 Flash NextMost efficient long contextChina180B256K tokensQwen Community Licensetext, imageAug 2026
18OpenAI logogpt-oss-120bBest on a single GPUUS117B / 5.1B128K tokensApache 2.0textAug 2025Transaction monitoring

Parameters show total / active per token for mixture-of-experts models.

Open source vs frontier models

Frontier models lead on the hardest tasks. Open models win on cost and control.

DimensionOpen source modelsFrontier models
CostA fraction of frontier pricingPremium on top tiers
Data controlRuns in your cloudProvider's data policy
CustomizationFine-tune the weightsPrompting only
Peak capabilityClose behindStill leads on the hardest tasks
HostingSelf-host, Bedrock, or Vertex AIProvider API only

Run open models inside your own cloud

Pick a model per agent or per skill. Evals tell you when a cheaper one is good enough.

Your cloud

Bedrock, Vertex AI, or your own vLLM or SGLang endpoint. Card data stays put.

Any harness

OpenCode, Cline, or Claude Code with an Anthropic-compatible API.

Routed by evals

Move routine work to the smallest model that still passes.

Understanding open source AI models

What are open source AI models?

Language models whose weights you can download, run, and usually modify. Host them yourself, on Bedrock or Vertex AI, or inside Runtime.

Open weight vs open source

Most are open weight: the weights are public, the training data is not. MIT, Apache 2.0, and OpenMDW are permissive. Custom licenses can add conditions.

Why mixture-of-experts is cheap

Each token uses a small slice of the network, so an 800B model may only run 40B parameters per token.

How to choose

Match context to your inputs, check the license, then test a few models on your real cases.

Open source AI models FAQ

What is the best open source AI model right now?

+

As of November 11, 2026, DeepSeek V4.1 Flash from DeepSeek. DeepSeek's October update to V4.1 Flash beats Kimi K3 and GLM-5.3 on every agent and coding benchmark in its report, with only 8B active parameters.

Which open source model is best for coding?

+

Kimi K2.7 Code from Moonshot AI. A coding-focused Kimi tuned for long-horizon software engineering.

What is the best open source model made in the US?

+

Muse Spark Open from Meta ranks highest. The other US models in our list are Nemotron 3.5 Super VL (NVIDIA), Muse Glimmer (Meta), Nemotron 3.5 Bolt (NVIDIA), gpt-oss-120b (OpenAI).

What is the best open source model made in Europe?

+

Mistral Large 4 from Mistral AI. Mistral's largest model, a 1T-parameter multimodal MoE that ties GLM-5.3 on Artificial Analysis's coding agent index with fewer tokens per task.

Which open source models can I use commercially?

+

MIT, Apache 2.0, and OpenMDW allow commercial use and fine-tuning: DeepSeek V4.1 Flash, DeepSeek V4 Pro, GLM-5.3 Flash, Nemotron 3.5 Super VL, Qwen3.8 27B, Muse Glimmer, Nemotron 3.5 Bolt, gpt-oss-120b. The rest use custom licenses, so read the terms first.

Are open source models as good as frontier models?

+

Frontier models still lead on the hardest tasks. Open models are close on most agent benchmarks and win on cost, privacy, and control.

What is the difference between open source and open weight?

+

Open weight means the trained weights are public. Fully open source also publishes the training data and code. Most "open source" models are open weight, so check the license.

Which open source model is best to self-host?

+

Qwen3.8 27B. A compact dense vision-language model under Apache 2.0. For one 80GB GPU, try gpt-oss-120b.

Can I run these open source models in Runtime?

+

Yes. Runtime is model-neutral. Serve a model from Bedrock, Vertex AI, or your own endpoint, and use evals to move routine work to cheaper models.

Run open source models in Runtime

Pick a model and put an agent to work in your own cloud.