Introducing the SEA-LION v4.5 Suite – Agentic Power and Speed
We are glad to introduce the SEA-LION v4.5 family, which adds state-of-the-art models that work well as agents but also understand Southeast Asian (SEA) languages, cultures, and local knowledge to the SEA-LION family of models.
SEA-LION v4.5 models benefit from new post-training techniques: knowledge distillation, targeted supervised fine-tuning, and smarter model merging. With these methods, we have taken SOTA models from Qwen 3.6 and Gemma 4 and adapted them for the diverse regional and local contexts of SEA. We also trained our own speculative decoder to turbo charge our large Qwen model, for even greater efficiency (i.e. up to 5X).
SEA-LION v4.5
- Custom distillation from Qwen3.5-397B-A17B and Gemma-4-31B-IT on an updated aisingapore/SEA-Instruct-2602
- Focus on multilingual and multicultural fluency across English and key SEA languages including Burmese, Indonesian, Filipino, Malay, Tamil, Thai, and Vietnamese.
- This set of models is designed to give developers and enterprises immediate access to high-capacity reasoning and agentic capabilities in English and SEA regional contexts even in resource-constrained environments.
1. Gemma-SEA-LION-v4.5-E2B-IT (Small and Efficient Agentic Base)
Built upon Gemma-4-E2B-IT, this highly streamlined model effectively operates at the size of a 2B model to maximise parameter efficiency and is optimised for resource-constrained environments. It maintains exceptional performance-to-size efficiency, making it highly effective for automated agentic tool use in localised SEA contexts with minimal memory overhead.
2. Qwen-SEA-LION-v4.5-27B-IT (Powerful Agent with Low Latency)
Built upon Qwen3.6-27B, this model preserves the advanced instruction following and multimodal agentic capabilities of its base while operating with deep linguistic and cultural understanding across SEA languages including Burmese, Indonesian, Filipino, Malay, Tamil, Thai, and Vietnamese. It features a native 262K context window, making it highly suitable for complex and agentic reasoning, document processing and enterprise RAG workflows.
3. Qwen-SEA-LION-v4.5-27B-IT-SpecDecoder (High-Throughput Booster)
To address the demands of latency-critical production environments, we are releasing a speed-optimising drafter model to complement Qwen-SEA-LION-v4.5-27B-IT, built upon z-lab’s Qwen3.6-27B-DFlash. Unlike conventional Multi Token Prediction (MTP), which drafts tokens sequentially, Qwen-SEA-LION-v4.5-27B-SpecDecoder is based on a block diffusion model that drafts blocks of tokens in one go, resulting in a much more significant speedup without sacrificing quality of output. Qwen-SEA-LION-v4.5-27B-IT-SpecDecoder works in perfect tandem with Qwen-SEA-LION-v4.5-27B-IT to dramatically increase tokens-per-second and reduce overall operational costs.
| Dataset | Baseline aisingapore/Qwen-SEA-LION-v4.5-27B-IT | w/ MTP | w/ DFlash z-lab/Qwen3.6-27B-Dflash | w/ SpecDecoder aisingapore/Qwen-SEA-LION-v4.5-27B-IT-SpecDecoder |
|---|---|---|---|---|
| gsm8k | 64.55 tok/s | 149.19 tok/s 2.31x | 283.67 tok/s 4.39x | 324.32 tok/s 5.02x |
| math500 | 65.91 tok/s | 153.19 tok/s 2.32x | 306.96 tok/s 4.66x | 335.22 tok/s 5.09x |
| humaneval | 66.03 tok/s | 155.52 tok/s 2.36x | 374.13 tok/s 5.67x | 397.11 tok/s 6.01x |
| mbpp | 66.44 tok/s | 148.69 tok/s 2.24x | 235.37 tok/s 3.54x | 260.46 tok/s 3.92x |
| mt-bench | 66.40 tok/s | 136.89 tok/s 2.06x | 153.81 tok/s 2.32x | 163.37 tok/s 2.46x |
| Tagalog | 66.45 tok/s | 117.06 tok/s 1.76x | 79.91 tok/s 1.20x | 164.53 tok/s 2.48x |
| Burmese | 66.46 tok/s | 134.61 tok/s 2.03x | 85.47 tok/s 1.29x | 266.73 tok/s 4.01x |
| Tamil | 66.48 tok/s | 111.99 tok/s 1.68x | 76.47 tok/s 1.15x | 175.82 tok/s 2.64x |
| Indonesian | 66.47 tok/s | 119.68 tok/s 1.80x | 90.30 tok/s 1.36x | 134.24 tok/s 2.02x |
| Vietnamese | 66.48 tok/s | 112.04 tok/s 1.69x | 87.51 tok/s 1.32x | 133.95 tok/s 2.01x |
| Thai | 66.28 tok/s | 104.81 tok/s 1.58x | 75.37 tok/s 1.14x | 118.49 tok/s 1.79x |
| Chinese | 66.46 tok/s | 130.32 tok/s 1.96x | 113.88 tok/s 1.71x | 133.43 tok/s 2.01x |
| Malay | 66.44 tok/s | 127.31 tok/s 1.92x | 106.24 tok/s 1.60x | 169.91 tok/s 2.56x |
*All the parameters setting follows default. num_speculative_tokens is 16
Speedup for Malay with SEA-LION-v4.5-27B-IT-SpecDecoder (right) compared with baseline (left)
Speedup for Burmese with SEA-LION-v4.5-27B-IT-SpecDecoder (bottom) compared with baseline (top)
Performance and Evaluation
On the SEA-HELM leaderboard, SEA-LION-v4.5 demonstrates improvement in multilingual and multicultural capabilities over their counterparts:
Screenshot from SEA-HELM leaderboard
*We are constantly updating the leaderboard – more to come very soon!
Claw-Eval is an open benchmark that evaluates autonomous agents on 300 tasks across 3 splits and 9 categories, each task with human-verified rubrics. Tasks include (i) multi-turn conversational tasks with simulated user personas asking for clarification and advice, (ii) multimodal tasks testing perception and generation, as well as (iii) general tasks covering core agent tasks across communication, finance, ops, productivity and so on.
SEA-LION-v4.5 delivers strong multilingual and multicultural performance while retaining its agentic capabilities:
| Model | Pass^3 |
|---|---|
| aisingapore/Gemma-SEA-LION-v4.5-E2B-IT | 9/107 (8%) |
| google/gemma-4-E2B-it | 11/107 (10%) |
| google/gemma-4-E4B-it | 7/107 (7%) |
| google/gemma-4-31B-it | 22/107 (21%) |
| aisingapore/Qwen-SEA-LION-v4.5-27B-IT | 48/107 (45%) |
| Qwen/Qwen3.6-27B | 47/107 (44%) |
| Qwen/Qwen3.5-27B | 42/107 (39%) |
Scores for Claw-Eval with thinking enabled
| Model | Pass^3 |
|---|---|
| aisingapore/Gemma-SEA-LION-v4.5-E2B-IT | 9/107 (8%) |
| google/gemma-4-E2B-it | 7/107 (7%) |
| google/gemma-4-E4B-it | 7/107 (7%) |
| google/gemma-4-31B-it | 23/107 (21%) |
| aisingapore/Qwen-SEA-LION-v4.5-27B-IT | 43/107 (40%) |
| Qwen/Qwen3.6-27B | 32/107 (30%) |
| Qwen/Qwen3.5-27B | 34/107 (32%) |
Scores for Claw-Eval with thinking disabled
Get Started Today
The model weights, detailed model cards, and starter code are now available on our Hugging Face hub:
- Explore the Collection: Hugging Face / AI Singapore
- Review the Benchmark Scores: SEA-LION Leaderboard
If you are building with the SEA-LION v4.5 suite, we would love to hear from you. For deployment inquiries or technical feedback, contact our team at sealion@aisingapore.org.