Motif-Technologies/Motif-3

---

1. Model Introduction

Motif 3 is a large-scale, decoder-only Mixture-of-Experts (MoE) language model with 314 billion total parameters and 13.2 billion parameters activated per token. It is built from the ground up by Motif Technologies following a fully in-house, proprietary design.

Motif 3 is built around Grouped Differential Latent Attention (GDLA), which integrates grouped differential attention with the compressed key–value representation of Multi-head Latent Attention. The architecture further incorporates modified manifold-constrained hyper-connections (mHC), Expert-Specific PolyNorm activations, and a Multi-Token Prediction (MTP) head to improve optimization stability, expert specialization, and inference efficiency.

The model is pretrained on approximately 12.5 trillion tokens spanning web documents, STEM, code, mathematics, multilingual content, and domain-specialized corpora, with additional emphasis on Korean, reasoning-intensive, legal, and financial data. Post-training combines general supervised fine-tuning, six RL-trained specialist teachers, a software-engineering teacher, and Multi-teacher On-Policy Distillation (MOPD) into a single unified model.

Key Features

• 🧠 Fine-grained sparse MoE — 384 routed experts with only 8 activated per token (plus 1 shared expert), providing a large expert pool at limited per-token compute.
• 📏 Native 256K context (262,144 tokens), trained with window-aware context parallelism.
• ⚙️ Novel architecture — GDLA attention, Expert-Specific PolyNorm, modified mHC, and a built-in MTP head enabling self-speculative decoding.
• 🌐 Multilingual & general-purpose, with a strong bytes-per-token tokenizer for English, Korean, code, and math.
• 🎯 Agentic strengths — particularly strong on long-horizon agentic tool use and terminal-based problem solving, with calibrated abstention on hallucination-sensitive evaluations.

2. Model Summary

Architecture	Mixture-of-Experts (MoE), decoder-only
Total Parameters	~314B
Activated Parameters	~13.2B / token
Number of Layers	53 (2 dense + 51 MoE)
Hidden Dimension	4096
Dense FFN Intermediate	12,288 (first 2 layers)
Attention	Grouped Differential Latent Attention (GDLA) with gated output
Query / KV Heads	80 / 16
Routed Experts	384 (top-8)
Shared Experts	1
Activation	Expert-Specific PolyNorm
Residual	Modified manifold-constrained hyper-connections (mHC)
MTP Head	1 layer (self-speculative decoding)
Context Length	262,144 (256K)
Vocabulary Size	220,160
Pretraining Tokens	~12.5T
Tensor Type	bfloat16

3. Evaluation Results

For contextual comparison, Motif 3 is compared with strong open-weight models using scores reported on the corresponding benchmark leaderboards. All Motif 3 evaluations were performed with sampling temperature = 1.0, top-p = 0.95, and a maximum sequence length of 262,144 tokens.
(*: public dataset only)

Benchmark	Motif 3 314B-A13B	MiniMax-3 428B-A23B	GLM-5.1 744B-A40B	Kimi-K2.6 1T-A32B	Qwen-3.7 max	DS-v4-Pro 1.6T-A49B
Agentic
GDPVal v2	38.7	44.4	37.8	34.4	39.0	40.2
τ²-Bench Telecom	94.7	88.9	97.7	95.9	94.7	96.2
τ³-Banking	35.3	15.3	13.6	23.3	12.0	30.1
ITBench*	51.5	—	40.3	31.2	42.5	38.3
Coding
SWE-Bench Verified	76.2	75.0	76.4	76.2	80.4	77.4
Terminal-Bench 2.1	74.9	65.2	61.8	65.9	75.0	64.0
SciCode	40.6	45.4	43.8	53.5	53.5	50.0
Reasoning & Knowledge
IMOAnswerBench	83.2	—	83.8	81.8	90.0	89.8
Apex-Shortlist	75.5	—	71.1	77.4	44.5	85.8
GPQA Diamond	83.4	92.9	86.8	91.1	92.4	88.8
HLE	37.0	39.0	30.1	37.5	41.4	37.5
CritPt	6.6	3.7	4.6	8.0	11.4	12.9
OmniScience — Accuracy	30.1	16.7	23.7	32.6	31.0	42.9
OmniScience — Non-Hallucination	71.6	81.6	70.1	59.5	74	5.9
Long Context & Instruction Following
AA-LCR	72.3	80.3	68.0	76.7	75.0	70.0
IFBench	78.2	82.9	76.3	76.0	79.1	76.5

_{Comparison scores are taken from the corresponding benchmark leaderboards.}

Motif 3 performs particularly well on agentic and tool-oriented benchmarks, while maintaining competitive performance across coding, mathematical reasoning, and general knowledge. On AA-Omniscience it pairs its accuracy with one of the highest non-hallucination scores, indicating a favorable balance between answering correctly and abstaining when unsupported.

4. Architecture

Motif 3 is a fully in-house design and introduces several custom components (full details in the technical report):

• Grouped Differential Latent Attention (GDLA) — integrates grouped differential attention (asymmetric signal/noise heads with a token-dependent differential coefficient) with the compressed KV latent of Multi-head Latent Attention, plus a query-dependent output gate. Retains the expressive attention dynamics of differential attention while substantially reducing KV-cache requirements.
• Expert-Specific PolyNorm — replaces the SiLU gate with a learned polynomial normalization whose coefficients are learned independently per expert, reducing activation outliers while allowing each expert to specialize.
• Modified manifold-constrained hyper-connections (mHC) — replaces conventional residual additions with a doubly-stochastic (Birkhoff-polytope) mixing of 4 parallel residual streams; the post-mapping multiplier is annealed from 2 → 1 during pretraining to limit activation-outlier accumulation.
• Multi-Token Prediction (MTP) — a 1-layer MTP head (DeepSeek-V3 style) used as an auxiliary pretraining objective and enabling self-speculative decoding at inference.

5. Deployment — vLLM (Recommended)

• Tested on B200 and H200 GPUs.
• The model ships with a built-in MTP head (num_nextn_predict_layers=1), so it supports self-speculative decoding — add --speculative-config as shown below (num_speculative_tokens: 1 is optimal for this model).
• Supports online block-fp8 quantization with --quantization modelopt_blockfp8
• If you encounter any issues, please open an HF issue.

Looking for a smaller footprint? An NVFP4-quantized checkpoint is available at Motif-Technologies/Motif-3-NVFP4.

• Docker image: ghcr.io/motiftechnologies/vllm:v0.20.2-motif3.rc3

5.1 H200

vllm serve "Motif-Technologies/Motif-3" \
    --trust-remote-code \
    --quantization modelopt_blockfp8 \
    --speculative-config '{"model": "Motif-Technologies/Motif-3", "num_speculative_tokens": 1}' \
    --tensor-parallel-size 1 \
    --data-parallel-size 8 \
    --data-parallel-size-local 8 \
    --enable-expert-parallel \
    --enable-auto-tool-choice \
    --tool-call-parser motif \
    --reasoning-parser motif \
    --dtype bfloat16 \
    --gpu-memory-utilization 0.85 \
    --max-model-len 262144 \
    --block-size 128 \
    --attention-backend FLASH_ATTN_MLA \
    --host 0.0.0.0 --port 8080

5.2 B200

vllm serve "Motif-Technologies/Motif-3" \
    --trust-remote-code \
    --quantization modelopt_blockfp8 \
    --speculative-config '{"model": "Motif-Technologies/Motif-3", "num_speculative_tokens": 1}' \
    --tensor-parallel-size 1 \
    --data-parallel-size 8 \
    --data-parallel-size-local 8 \
    --enable-expert-parallel \
    --enable-auto-tool-choice \
    --tool-call-parser motif \
    --reasoning-parser motif \
    --dtype bfloat16 \
    --gpu-memory-utilization 0.85 \
    --max-model-len 262144 \
    --block-size 128 \
    --host 0.0.0.0 --port 8080

6. Access

This model is openly available — anyone can download the weights, no access request required.

7. License

This model is released under the MIT License. See the LICENSE file for details.

---

© Motif Technologies. All rights reserved.