MiniMax-M3
MiniMax · Multimodal reasoning model
- Parameters
- 427.04B
- License
- MiniMax Community License
- BF16 weights
- 854.08 GB (weight-only estimate)
OPEN-WEIGHT AGENT INTELLIGENCE
Models for AI agents.
Discover open-weight model components for AI agents, from reasoning cores to retrieval, decision and media tools. Explore 12 Leading and 19 Notable models with source-linked facts—not invented agent-readiness scores.
31Curated models
12Leading
19Notable
0Assumed agent benchmarks
START WITH THE TASK
Six paths to a useful agent.
Agent architecture determines which model component matters. A retrieval encoder or image generator may support an agent without acting as the planning model.
EXPLORE USE CASETool CallingTool calling is a system behavior, not a guarantee conferred by downloadable weights. Assess the model's tool-use training, chat template, serving parser, output validity and multi-step task success separately.Explore guide →EXPLORE USE CASECodingA useful coding agent must inspect a repository, produce applicable patches and pass tests under controlled permissions. A model's coding-oriented category alone cannot establish autonomous software-engineering performance.Explore guide →EXPLORE USE CASEResearchResearch-agent quality depends on retrieval, source selection, faithful citations and multi-step execution, not just the language model. Model-only reasoning benchmarks cannot substitute for end-to-end research evaluations.Explore guide →EXPLORE USE CASEDocuments & RAGA document agent normally combines extraction, indexed retrieval, a generation model and permissions. Embedding models support retrieval; they do not replace the agent's reasoning model.Explore guide →EXPLORE USE CASEComputer UseComputer-use capability requires a model or system to interpret screen state, plan actions and control an interface. Image-input support is not proof that a model can navigate a desktop reliably.Explore guide →EXPLORE USE CASEPrivate & LocalOpen-weight checkpoints can make self-hosted agents possible, but a private system requires local inference, tools, retrieval, logs and secure operations. Model licensing and hardware feasibility must be checked per artifact.Explore guide →
AGENT INTELLIGENCE 2.0
Trace the capability. Test the configuration.
Move beyond generic tags with 9 reviewed publisher capability claims, 7 attributed benchmark observations and 7 runtime references. These are not independently measured agent results.
MODEL EXPLORER
Make the shortlist explain itself.
All 31 Leading and Notable records are shown below. Filters narrow a research shortlist; they do not certify fitness for any task. Hardware values are mathematical weight-only estimates, not measured VRAM.
31 curated model records · results are not ranked
K2-Horizon-375B-A23B
Institute of Foundation Models (IFM) · Reasoning model
- Parameters
- 379.167B
- License
- apache-2.0
- BF16 weights
- 758.33 GB (weight-only estimate)
MiMo-V2.6-Pro-RL
Xiaomi MiMo · Omnimodal reasoning model
- Parameters
- 1024.22B
- License
- mit
- BF16 weights
- 2,048.43 GB (weight-only estimate)
GLM-5.3
Z.ai · Text reasoning model
- Parameters
- 753.33B
- License
- other
- BF16 weights
- 1,506.66 GB (weight-only estimate)
Kimi-K3
Moonshot AI · Multimodal reasoning model
- Parameters
- 2779.93B
- License
- other
- BF16 weights
- 5,559.86 GB (weight-only estimate)
GLM-5.3-Flash
Z.ai · Multimodal reasoning model
- Parameters
- 321.323B
- License
- mit
- BF16 weights
- 642.65 GB (weight-only estimate)
DeepSeek-V4.1-Flash
DeepSeek · Text reasoning model
- Parameters
- 763.205B
- License
- mit
- BF16 weights
- 1,526.41 GB (weight-only estimate)
Qwen3.8-27B
Qwen · Multimodal reasoning model
- Parameters
- 28B
- License
- apache-2.0
- BF16 weights
- 55.56 GB (weight-only estimate)
Inkling
thinkingmachines · Text reasoning model
- Parameters
- 952.378B
- License
- apache-2.0
- BF16 weights
- 1,904.76 GB (weight-only estimate)
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
NVIDIA · Text reasoning model
- Parameters
- 560.525B
- License
- other
- BF16 weights
- 1,121.05 GB (weight-only estimate)
Muse-Glimmer-30B
meta-models · Text reasoning model
- Parameters
- 29.7766B
- License
- apache-2.0
- BF16 weights
- 59.55 GB (weight-only estimate)
gpt-oss-120b
OpenAI · Text reasoning model
- Parameters
- 117B total / 5.1B active
- License
- apache-2.0
- BF16 weights
- 233.66 GB (weight-only estimate)
gemma-4-26B-A4B-it
Google · Multimodal reasoning model
- Parameters
- 25.8059B
- License
- apache-2.0
- BF16 weights
- 51.61 GB (weight-only estimate)
Mistral Small 4 119B A6B
Mistral AI · Multimodal reasoning model
- Parameters
- 119B total / 6.5B active
- License
- apache-2.0
- BF16 weights
- 238.80 GB (weight-only estimate)
Qwen-Image-2.1
Qwen · Image generation or editing
- Parameters
- 7B
- License
- other
- BF16 weights
- 14.23 GB (weight-only estimate)
FLUX.2-dev
Black Forest Labs · Image generation or editing
- Parameters
- 32.2233B
- License
- other
- BF16 weights
- 64.45 GB (weight-only estimate)
MiniMax-H3
MiniMax · Video generation
- Parameters
- 33.123B
- License
- other
- BF16 weights
- 66.25 GB (weight-only estimate)
LTX-2.5
Lightricks · Video generation
- Parameters
- Pending verification
- License
- other
- BF16 weights
- Not calculated
Voxtral-4B-TTS-2603
Mistral AI · Speech generation
- Parameters
- 4B
- License
- cc-by-nc-4.0
- BF16 weights
- 8.00 GB (weight-only estimate)
Olmo-3-7B-Think
AllenAI · Open research reasoning model
- Parameters
- 7.29801B
- License
- apache-2.0
- BF16 weights
- 14.60 GB (weight-only estimate)
K2-Horizon-3.7B
Institute of Foundation Models (IFM) · Compact reasoning model
- Parameters
- 5.05826B
- License
- apache-2.0
- BF16 weights
- 10.12 GB (weight-only estimate)
Mellum2.1-12B-A2.5B-Thinking
JetBrains · Coding-oriented model
- Parameters
- 12.1499B
- License
- apache-2.0
- BF16 weights
- 24.30 GB (weight-only estimate)
K2-Horizon-MoVA-36B-A4B
Institute of Foundation Models (IFM) · Sparse reasoning model
- Parameters
- 37.4448B
- License
- apache-2.0
- BF16 weights
- 74.89 GB (weight-only estimate)
clef-omni
Cloudflare · Decision scorer
- Parameters
- 35.2598B
- License
- apache-2.0
- BF16 weights
- 70.52 GB (weight-only estimate)
pplx-embed-v2-late-0.6b
Perplexity · Retrieval embeddings
- Parameters
- 0.594322B
- License
- mit
- BF16 weights
- 1.19 GB (weight-only estimate)
pplx-embed-v2-late-9b
Perplexity · Retrieval embeddings
- Parameters
- 8.3927B
- License
- mit
- BF16 weights
- 16.79 GB (weight-only estimate)
Qwen-Image-2.1-Turbo
Qwen · Image generation or editing
- Parameters
- 7.11512B
- License
- other
- BF16 weights
- 14.23 GB (weight-only estimate)
Darwin-27B-ZTC
FINAL-Bench · Typed decision scorer
- Parameters
- 25.6246B
- License
- apache-2.0
- BF16 weights
- 51.25 GB (weight-only estimate)
Holo4-27B
H Company · Computer-use vision-language model
- Parameters
- 27.3567B
- License
- cc-by-nc-4.0
- BF16 weights
- 54.71 GB (weight-only estimate)
Granite 4.2 30B
IBM · Text reasoning model
- Parameters
- 29.2768B
- License
- apache-2.0
- BF16 weights
- 58.55 GB (weight-only estimate)
LightOnOCR-3-4B
LightOn AI · Document OCR & layout
- Parameters
- 4.53927B
- License
- apache-2.0
- BF16 weights
- 9.08 GB (weight-only estimate)
No model here is marked “agent ready.” Tool-call accuracy, GUI control and multi-step success are unknown unless independently assessed for the precise model-and-runtime combination.
FROM WEIGHTS TO SYSTEMS
Practical deployment, without assumptions.
OllamaRun a Local Agent Model with Ollama →llama.cppServe Local Agent Models with llama.cpp →vLLMDeploy Agent Models with vLLM →
A runtime supporting tool calls does not establish that a particular checkpoint works in that runtime. Verify weight format, chat template, parser, hardware and license first.
Questions about model selection
Are all 31 models complete AI agents?
No. The selection includes reasoning model candidates, retrieval/decision components and media models; an agent also needs orchestration, tools and safety controls.
Are Leading and Notable agent performance ratings?
No. These are existing curated catalog labels, not agent-task benchmark scores.