DeepSeek V4.1 Flash
DeepSeek China
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.
- Context
- 1M tokens
- Params
- 552B · 8B active
- Input
- text, image
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
Ranked for agent work: tool use, context, license, and cost at volume.
DeepSeek China
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.
Moonshot AI China
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.
Z.ai China
Z.ai's strongest open-weight model, post-trained for complex coding and long-horizon agent work.
Moonshot AI China
The first open 3T-class model, natively multimodal with a 1M-token context.
MiniMax China
MiniMax's October update to M3, built for coding, agent work, computer use, and security, with the top open-weight BankerToolBench score in MiniMax's report.
DeepSeek China
The official DeepSeek V4 Pro release, with stronger agentic skills and an MIT license.
Meta US
Meta's open-weight Muse Spark, built for coding and agent work with a 1M-token context.
Mistral AI France
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.
Z.ai China
A natively multimodal GLM that Z.ai says beats GLM-5.2 at one-tenth the price.
Moonshot AI China
A coding-focused Kimi tuned for long-horizon software engineering.
NVIDIA US
NVIDIA's vision-language agent model, with 12B active parameters and a 1M-token context.
Qwen (Alibaba) China
The first Qwen-Max-class model released with open weights.
MiniMax China
A natively multimodal model with sparse attention built for million-token contexts.
Qwen (Alibaba) China
A compact dense vision-language model under Apache 2.0.
Meta US
A 30B agent model distilled from Muse Spark that runs on one 24GB to 32GB GPU or a Mac.
NVIDIA US
A 74B model with 9B active that NVIDIA also ships in a 2-bit build sized for 24GB.
Qwen (Alibaba) China
An experimental preview of the architecture that will underpin Qwen4.
OpenAI US
OpenAI's open-weight reasoning model, small enough to run on a single 80GB GPU.
Benchmark scores are reported by each model provider on its Hugging Face model card. They are not independently verified by Runtime.
Filter by license, input, or origin. Sort by context or release date.
| Origin | Parameters | License | Input | Good for | ||||
|---|---|---|---|---|---|---|---|---|
| 1 | China | 552B / 8B | 1M tokens | MIT | text, image | Oct 2026 | Transaction monitoringKYC and KYB checksPayments customer support | |
| 2 | China | 2.8T / 104B | 1M tokens | Kimi K3.1 Pro License | text, image | Nov 2026 | Chargebacks and disputes | |
| 3 | China | 753B | 1M tokens | GLM-5.3 License | text | Aug 2026 | Payment reconciliationChargebacks and disputesMerchant underwriting | |
| 4 | China | 2.8T | 1M tokens | Kimi K3 License | text, image, video | Jun 2026 | Partner-bank RFIsMerchant underwriting | |
| 5 | China | 428B / 23B | 1M tokens | MiniMax Community License | text, image, video | Oct 2026 | Partner-bank RFIsChargebacks and disputes | |
| 6 | China | 1.6T | 1M tokens | MIT | text | Aug 2026 | Payment reconciliationMerchant underwritingAML and compliance investigations | |
| 7 | US | 816B / 42B | 1M tokens | Muse Model License | text, image, audio, video | Oct 2026 | ||
| 8 | France | 1T / 52B | 1M tokens | Mistral Ecosystem Open-Weight License | text, image | Sep 2026 | ||
| 9 | China | 320B / 18B | 1M tokens | MIT | text, image | Aug 2026 | Partner-bank RFIsTransaction monitoringKYC and KYB checks | |
| 10 | China | 1T / 32B | 256K tokens | Modified MIT | text, image | Jun 2026 | ||
| 11 | US | 120B / 12B | 1M tokens | OpenMDW-1.1 | text, image, video | Oct 2026 | ||
| 12 | China | 2.4T / 95B | 256K tokens | Qwen3.8-Max License | text | Aug 2026 | AML and compliance investigationsPayments customer support | |
| 13 | China | 428B / 23B | 1M tokens | MiniMax Community License | text, image, video | Jun 2026 | ||
| 14 | China | 27B | 256K tokens | Apache 2.0 | text, image, video | Aug 2026 | Payment reconciliationKYC and KYB checks | |
| 15 | US | 30B | 128K tokens | Apache 2.0 | text, image | Aug 2026 | ||
| 16 | US | 74B / 9B | 256K tokens | OpenMDW-1.1 | text, image, video | Oct 2026 | ||
| 17 | China | 180B | 256K tokens | Qwen Community License | text, image | Aug 2026 | ||
| 18 | US | 117B / 5.1B | 128K tokens | Apache 2.0 | text | Aug 2025 | Transaction monitoring |
Parameters show total / active per token for mixture-of-experts models.
Weights run wherever you host them. Origin still comes up in vendor and model risk reviews, so here is where each provider is based.
Our picks for the work payment teams hand to agents.
Frontier models lead on the hardest tasks. Open models win on cost and control.
| Dimension | Open source models | Frontier models |
|---|---|---|
| Cost | A fraction of frontier pricing | Premium on top tiers |
| Data control | Runs in your cloud | Provider's data policy |
| Customization | Fine-tune the weights | Prompting only |
| Peak capability | Close behind | Still leads on the hardest tasks |
| Hosting | Self-host, Bedrock, or Vertex AI | Provider API only |
Pick a model per agent or per skill. Evals tell you when a cheaper one is good enough.
Bedrock, Vertex AI, or your own vLLM or SGLang endpoint. Card data stays put.
OpenCode, Cline, or Claude Code with an Anthropic-compatible API.
Move routine work to the smallest model that still passes.
Language models whose weights you can download, run, and usually modify. Host them yourself, on Bedrock or Vertex AI, or inside Runtime.
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.
Each token uses a small slice of the network, so an 800B model may only run 40B parameters per token.
Match context to your inputs, check the license, then test a few models on your real cases.
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.
Kimi K2.7 Code from Moonshot AI. A coding-focused Kimi tuned for long-horizon software engineering.
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).
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.
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.
Frontier models still lead on the hardest tasks. Open models are close on most agent benchmarks and win on cost, privacy, and control.
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.
Qwen3.8 27B. A compact dense vision-language model under Apache 2.0. For one 80GB GPU, try gpt-oss-120b.
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.
Pick a model and put an agent to work in your own cloud.