The best open source AI models, ranked and compared

GLM-5.3 keeps the GLM-5.2 base model and puts every gain into post-training. Z.ai reports open-source state of the art on Terminal Bench 3.0 and Agents' Last Exam, and a 50% gain over GLM-5.2 on its in-house code benchmark.

• 1M-token context
• Top open-weight scores on agentic benchmarks
• Built for long, multi-step tool use

Kimi K3 is a 2.8T-parameter mixture-of-experts model that activates 16 of 896 experts per token. Moonshot built it for long-horizon coding and agentic knowledge work, and it reads text, images, and video in the same model.

• 2.8T parameters
• Text, image, and video input
• 1M-token context

DeepSeek-V4-Pro-0813 supersedes the V4 Pro preview with much better agentic performance in production settings. It adds a DSpark speculative decoding module and is released under MIT, so commercial use and fine-tuning are straightforward.

• MIT license
• 1M-token context
• Strong reasoning and tool use

GLM-5.3 Flash starts from a new base model with hybrid sparse and linear attention, which cuts long-context serving costs. With 18B active parameters it is cheap to run, and Z.ai reports it approaching Claude Opus 4.8 on coding and agentic benchmarks.

• 18B active parameters
• MIT license
• Image input and 1M context

DeepSeek V4.1 Flash uses a causal encoder-decoder design that lets the decoder reuse a compressed KV cache, about 4x smaller than V4 Flash. It reads images and text, supports a 1M-token context, and lets you dial reasoning effort up or down.

• 8B active parameters
• About 4x smaller KV cache than V4 Flash
• MIT license

Kimi K2.7 Code builds on Kimi K2.6 with large gains on real-world, multi-step coding tasks. Moonshot reports about 30% fewer thinking tokens than K2.6, which lowers cost on long runs.

• Built for long-horizon coding
• About 30% fewer thinking tokens than K2.6
• 256K context

Qwen3.8 2.4T-A95B is the open-weight model behind Qwen3.8-Max. Alibaba reports large gains in coding, professional work, and long-horizon agent tasks over Qwen3.5 and 3.6, with tunable reasoning effort.

• 2.4T total, 95B active
• Qwen-Max class, open weights
• Tunable reasoning effort

MiniMax M3 trains on text, images, and video from the first step. Its MiniMax Sparse Attention gives 9x faster prefill and 15x faster decode than M2 at 1M context, and it has enabled, adaptive, and disabled reasoning modes.

• Text, image, and video input
• 9x prefill and 15x decode speedup vs M2 at 1M context
• Adaptive reasoning mode

Qwen3.8 27B brings the Qwen3.8 generation to a deployment-friendly dense model. It understands images and video, thinking can be switched off per request, and the Apache 2.0 license makes it simple to run in your own cloud.

• Apache 2.0 license
• Small enough to self-host
• Image and video input

Qwen3.8 Flash Next pairs Gated DeltaNet with Qwen Sparse Attention to cut long-context latency, and adds n-gram embeddings for cheaper parameter scaling. Alibaba positions it as a preview of the Qwen4 architecture.

• Hybrid attention for low long-context latency
• Image input
• Preview of the Qwen4 architecture

gpt-oss-120b is OpenAI's open-weight reasoning model under Apache 2.0. It activates 5.1B parameters per token and fits on a single 80GB GPU, which makes it a common choice for private, low-cost deployments.

• Apache 2.0 license
• Runs on one 80GB GPU
• Configurable reasoning effort