Motif 3
Motif Technologies' 314B mixture of experts with only 13.2B active and a 256K context. MIT license, built for agents and terminal work.
In short. Motif 3 is a large AI model from the South Korean company Motif Technologies. It is built for agent work such as using tools and running terminal tasks, and it only activates about 4% of its parameters for each word. The MIT license lets companies use it commercially with almost no conditions.
Summary
Motif Technologies published the final Motif 3 weights on Hugging Face in August 2026, alongside a technical report dated 10 August. Sources differ by a few days on the exact release day, so this site lists the month only. Motif 3 is a text-only mixture-of-experts model with 314 billion parameters and 13.2 billion active per token. It has 53 layers (2 dense and 51 mixture-of-experts layers), 384 routed experts with 8 active plus one shared expert, and a 256K context. The design is Motif's own: Grouped Differential Latent Attention, modified manifold-constrained hyper-connections, per-expert PolyNorm activations and a one-layer multi-token prediction head that allows self-speculative decoding. Motif trained on about 12.5 trillion tokens, with extra Korean, reasoning, legal and financial data, and its post-training distills six reinforcement-learning specialists and a software-engineering teacher into one model.
In Motif's table, which takes the other models' scores from public leaderboards, Motif 3 scores 74.9 on Terminal-Bench 2.1, against 65.9 for Kimi-K2.6, 64.0 for DeepSeek-V4-Pro and 61.8 for GLM-5.1, with Qwen-3.7 max level at 75.0, and 35.3 on τ³-Banking, the highest in the table. On SWE-Bench Verified it scores 76.2, behind DeepSeek-V4-Pro at 77.4. It is clearly behind on knowledge and math tests: 83.4 on GPQA Diamond against 86.8 to 92.9 for the others, and 83.2 on IMOAnswerBench against 89.8 for DeepSeek-V4-Pro. Motif's own runs are the Motif 3 column, so treat all numbers as developer-reported.
The license is the standard MIT license. Motif asks, without requiring it, that you mention Built with Motif 3 when you share your work. The BF16 weights are about 630 GB and the NVFP4 build about 187 GB. The card's serving commands use eight H200 or B200 GPUs with online FP8 quantization, through Motif's vLLM fork (based on v0.26.0) or its Docker image, and the model needs trust-remote-code. Motif also publishes an example repository with its architecture and distributed training framework.
Good for
- Agent work that mixes tool use and terminal tasks, where Motif's own table shows its best results
- Low-cost serving for its size, since only 13.2B of the 314B parameters are active per token
- Commercial products, since the MIT license has almost no conditions
- Teams that want to study or adapt an unusual architecture, with the training example repository available
Watch out for
- It is text-only, and the card names English and Korean as its languages
- It is behind the strongest models on knowledge and math tests such as GPQA Diamond and IMOAnswerBench
- It needs Motif's vLLM fork or Docker image and trust-remote-code, and the card's commands use eight H200 or B200 GPUs
- The BF16 weights are about 630 GB
- Benchmark numbers are Motif's own
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Motif 3 NVFP4 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Motif 3 (BF16) BF16 | 314B 13.2B active | 722 GB | 361 GB | 217 GB | Big-memory desktop |
| Motif 3 NVFP4 NVFP4 | 314B 13.2B active | 722 GB | 361 GB | 206 GB | Big-memory desktop |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
- Solar Open 2 vs Motif 3: which Korean agent model should you run?
- Best open-source models for coding and AI agents
- Open-weight AI models made in South Korea
- Korea's four sovereign AI models: open weights, three teams advance
- Giant open models now ship in 4-bit: what it means for memory planning
Frequently asked questions
- What is Motif 3?
- Motif 3 is a large AI model from the South Korean company Motif Technologies. It is built for agent work such as using tools and running terminal tasks, and it only activates about 4% of its parameters for each word. The MIT license lets companies use it commercially with almost no conditions.
- Can I use Motif 3 commercially?
- Yes. Motif 3 is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Motif 3 open source?
- Motif 3 is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Motif 3 on my own computer?
- Only on a machine with a lot of memory. The smallest variant, Motif 3 NVFP4, needs about 206 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
- When was Motif 3 released?
- Motif 3 was released by Motif Technologies on August 2026.
- How long a context window does Motif 3 have?
- Motif 3 supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Motif-3 model card (Hugging Face) huggingface.co
- Motif 3 technical report (arXiv 2608.09119) arxiv.org
- Motif Technologies models page motiftech.io
- Motif 3 MIT license (LICENSE file) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| A.X K2 SK Telecom | 688B 33B active | 256K | Apache 2.0 (Permissive) | 438 GB Big-memory desktop | |
| K-EXAONE 2.0 LG AI Research | 750B 37B active | 256K | Apache 2.0 (Permissive) | 518 GB GPU server | |
| Solar Open 2 Upstage | 250B 15B active | 1M | Upstage Solar (Conditional) | 157 GB Big-memory desktop | |
| GLM-5.2 Z.ai (Zhipu AI) | 753B 40B active | 1M | MIT (Permissive) | 520 GB GPU server |