--- library_name: transformers license: mit language: - en - zh - vi - id - th - fil - ta - ms - my license_link: https://huggingface.co/Qwen/Qwen3.6-27B/blob/main/LICENSE pipeline_tag: image-text-to-text base_model: - Qwen/Qwen3.6-27B --- ![Banner!](Qwen-SEA-LION-v4.5.png "Qwen-SEA-LION-v4.5-27B-IT") # Qwen-SEA-LION-v4.5-27B-IT _Last update: 2026-05-19_ **SEA-LION** is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned for the Southeast Asia (SEA) region. **Qwen-SEA-LION-v4.5-27B-IT** built upon the Qwen3.6-27B dense architecture, a 27-billion parameter model featuring a hybrid Linear and Full Attention design. To ensure deep domain adaptation, the model underwent distillation from Qwen3.5-397B-A17B on an updated aisingapore/SEA-Instruct-2602, instilling multilingual and multicultural fluency across English and key SEA languages including: Burmese, Indonesian, Filipino, Malay, Tamil, Thai, and Vietnamese. Qwen-SEA-LION-v4.5-27B-IT inherits the following features from Qwen3.6: - Context Window (262K): Large context window to enable strong reasoning capabilities. - Reasoning: Highly capable reasoning model, with configurable thinking modes. - Thinking Preservation: Retains historical reasoning context to streamline iterative development and reduce compute overhead. - Agentic Coding: High-precision handling of repository-level reasoning and frontend workflows. - Unified Vision-Language: Early fusion training delivers good performance across multimodal reasoning, coding, and visual tasks. ## Model Details ### Model Description SEA-LION stands for Southeast Asian Languages In One Network. We performed post-training in English and SEA languages on Qwen3.6-27B, a multimodal learning model using the Qwen3.6 architecture, to create Qwen-SEA-LION-v4.5-27B-IT. For tokenization, the model employs the default tokenizer used in Qwen3.6. - **Developed by:** AI Products Pillar, AI Singapore - **Funded by:** Singapore NRF - **Shared by:** AI Products Pillar, AI Singapore - **Model type:** Causal Language Model with Vision Encoder - **Training Stage:** Post-Training (Logit Distillation & Model Merging)) - **Context length:** 262k - **Language(s):** fine-tuned on Burmese, Indonesian, Filipino, Malay, Tamil, Thai, and Vietnamese - **License:** [MIT](https://mit-license.org/) - **Finetuned from model:** ### Model Sources - **Repository:** [SEA-LION v4.5 - an aisingapore Collection](https://huggingface.co/collections/aisingapore/sea-lion-v45) ## Uses ### Out-of-Scope Use The model has not been aligned for safety. Developers and users should perform their own safety fine-tuning and related security measures. In no event shall the authors be held liable for any claims, damages, or other liabilities arising from the use of the released weights and codes. ### Bias, Risks, and Limitations The model was not tested for robustness against adversarial prompting. It is important for users to be aware that our model exhibits certain limitations that warrant consideration. Like many LLMs, the model can hallucinate and occasionally generates irrelevant content, introducing fictional elements that are not grounded in the provided context. Users should also exercise caution in interpreting and validating the model's responses due to the potential inconsistencies. ## How to Get Started with the Model Use the code below to get started with the model with 🤗 Transformers libraries. ``` pip install "transformers>=4.57.0" accelerate ``` ```python # ============================================================ # TEXT-ONLY INFERENCE example # ============================================================ import torch from transformers import AutoTokenizer, AutoModelForCausalLM MODEL_ID = "aisingapore/Qwen-SEA-LION-v4.5-27B-IT" # ── Load tokenizer ── tokenizer = AutoTokenizer.from_pretrained(MODEL_ID) # ── Load model in bfloat16 across all available GPUs ── # attn_implementation="sdpa" is safer for the hybrid DeltaNet arch; # flash_attention_2 compatibility depends on your transformers version model = AutoModelForCausalLM.from_pretrained( MODEL_ID, torch_dtype=torch.bfloat16, device_map="auto", attn_implementation="sdpa", # use sdpa for DeltaNet hybrid layers ) # ── Message: text-only, same Malay query from original snippet ── messages = [ { "role": "user", "content": "Tolong carikan flat 4-bilik dekat Tampines, bajet bawah $500,000. " "Nak tahu juga berapa anggaran pinjaman bulanan." } ] # ── Apply chat template — text-only, thinking disabled ── # enable_thinking=False → instruct/non-thinking mode # Qwen3.6 does NOT support /no_think soft switch unlike Qwen3 text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False, # hard-disable CoT thinking blocks ) # ── Tokenize ── inputs = tokenizer(text, return_tensors="pt").to(model.device) # ── Generate — non-thinking mode params ── # presence_penalty=1.5 is important for Qwen3.6 non-thinking mode # to suppress repetition; not available in model.generate() directly, # so use do_sample=True with the temperature/top_p/top_k trio generated_ids = model.generate( **inputs, max_new_tokens=512, do_sample=True, temperature=0.7, # non-thinking instruct mode top_p=0.80, top_k=20, # Note: presence_penalty requires vLLM/SGLang for full effect; # in transformers use repetition_penalty as a proxy repetition_penalty=1.1, ) # ── Decode only newly generated tokens ── output_ids = generated_ids[0][inputs["input_ids"].shape[1]:] response = tokenizer.decode(output_ids, skip_special_tokens=True).strip() print(response) ``` #### Tool Calling example ```python # ============================================================ # TOOL CALLING (Transformers, local) # ============================================================ import torch from transformers import AutoTokenizer, AutoModelForCausalLM MODEL_ID = "aisingapore/Qwen-SEA-LION-v4.5-27B-IT" tokenizer = AutoTokenizer.from_pretrained(MODEL_ID) model = AutoModelForCausalLM.from_pretrained( MODEL_ID, torch_dtype=torch.bfloat16, device_map="auto", attn_implementation="sdpa", ) messages = [ { "role": "user", "content": "Tolong carikan flat 4-bilik dekat Tampines, bajet bawah $500,000. " "Nak tahu juga berapa anggaran pinjaman bulanan." } ] tools = [ { "type": "function", "function": { "name": "search_hdb_listings", "description": "Search for HDB flats available for sale", "parameters": { "type": "object", "properties": { "location": { "type": "string", "description": "Town or area name" }, "flat_type": { "type": "string", "description": "Flat type e.g. 3-room, 4-room, 5-room" }, "max_price": { "type": "number", "description": "Maximum price in SGD" } }, "required": ["location", "flat_type"] } } }, { "type": "function", "function": { "name": "calculate_mortgage", "description": "Calculate estimated monthly mortgage payment", "parameters": { "type": "object", "properties": { "loan_amount": { "type": "number", "description": "Loan amount in SGD" }, "interest_rate": { "type": "number", "description": "Annual interest rate as percentage" }, "loan_tenure_years": { "type": "integer", "description": "Loan period in years" } }, "required": ["loan_amount"] } } } ] # ============================================================ # apply_chat_template returns BatchEncoding with keys: # input_ids, attention_mask (and sometimes token_type_ids) # ============================================================ inputs = tokenizer.apply_chat_template( messages, tools=tools, return_tensors="pt", return_dict=True, # ← returns BatchEncoding dict with attention_mask add_generation_prompt=True, enable_thinking=False, # disable CoT for structured tool call output ).to(model.device) # ── Unpack BatchEncoding dict with ** — fixes the AttributeError ── generated_ids = model.generate( **inputs, # ← unpack: passes input_ids + attention_mask max_new_tokens=512, do_sample=False, ) # ── Decode only new tokens — slice off the prompt portion ── output_ids = generated_ids[0][inputs["input_ids"].shape[1]:] response = tokenizer.decode(output_ids, skip_special_tokens=True).strip() print(response) ``` **Agentic Example:** ```python # ============================================================ # NO-VLLM AGENTIC LOOP # ============================================================ import os import json import re import torch from transformers import AutoTokenizer, AutoModelForCausalLM from dotenv import load_dotenv load_dotenv() MODEL_ID = "aisingapore/Qwen-SEA-LION-v4.5-27B-IT" print("Loading tokenizer...") tokenizer = AutoTokenizer.from_pretrained( MODEL_ID, token=os.getenv("HF_TOKEN"), ) print("Loading model across GPUs...") model = AutoModelForCausalLM.from_pretrained( MODEL_ID, token=os.getenv("HF_TOKEN"), torch_dtype=torch.bfloat16, device_map="auto", attn_implementation="sdpa", ) device_info = getattr(model, "hf_device_map", None) or str(model.device) print(f"Model loaded. Device: {device_info}") TOOLS = [ { "type": "function", "function": { "name": "search_hdb_listings", "description": "Search for HDB flats available for sale", "parameters": { "type": "object", "properties": { "location": {"type": "string", "description": "Town or area name"}, "flat_type": {"type": "string", "description": "e.g. 4-room"}, "max_price": {"type": "number", "description": "Max price in SGD"}, }, "required": ["location", "flat_type"], }, }, }, { "type": "function", "function": { "name": "calculate_mortgage", "description": "Calculate estimated monthly mortgage payment", "parameters": { "type": "object", "properties": { "loan_amount": {"type": "number", "description": "Loan amount SGD"}, "interest_rate": {"type": "number", "description": "Annual rate %"}, "loan_tenure_years": {"type": "integer", "description": "Loan years"}, }, "required": ["loan_amount"], }, }, }, ] def execute_tool(name: str, arguments: dict) -> str: """Mock tool executor — replace with real API calls.""" if name == "search_hdb_listings": return json.dumps({ "listings": [ { "address": "Blk 472 Tampines St 43", "flat_type": arguments.get("flat_type"), "resale_price": 488000, "floor_area_sqm": 93, "remaining_lease": "67 years", }, { "address": "Blk 512 Tampines Ave 4", "flat_type": arguments.get("flat_type"), "resale_price": 475000, "floor_area_sqm": 89, "remaining_lease": "62 years", }, ] }) elif name == "calculate_mortgage": principal = arguments["loan_amount"] r = (arguments.get("interest_rate", 2.6) / 100) / 12 n = arguments.get("loan_tenure_years", 25) * 12 monthly = principal * (r * (1 + r) ** n) / ((1 + r) ** n - 1) return json.dumps({ "loan_amount": principal, "monthly_repayment_sgd": round(monthly, 2), }) return json.dumps({"error": f"Unknown tool: {name}"}) def generate_response(messages: list) -> str: """ Single model.generate() call. Returns the raw decoded string (may contain tool call JSON). """ # ── Render chat template to string first ── text = tokenizer.apply_chat_template( messages, tools=TOOLS, tokenize=False, add_generation_prompt=True, enable_thinking=False, # no blocks for tool calling ) # ── Tokenize separately ── inputs = tokenizer(text, return_tensors="pt").to(model.device) # ── Generate ── with torch.no_grad(): # saves memory during inference generated_ids = model.generate( **inputs, max_new_tokens=512, do_sample=False, # greedy for deterministic tool JSON ) # ── Decode new tokens only ── output_ids = generated_ids[0][inputs["input_ids"].shape[1]:] return tokenizer.decode(output_ids, skip_special_tokens=True).strip() def parse_tool_calls(response_text: str) -> list: """ Parse Hermes-style tool call JSON from model output. Qwen3.6 emits tool calls wrapped in ... tags. Returns list of {"name": ..., "arguments": {...}} dicts. Falls back to empty list if no tool calls found. """ import re tool_calls = [] # ── Match {...} blocks ── pattern = r"(.*?)" matches = re.findall(pattern, response_text, re.DOTALL) for match in matches: try: call = json.loads(match.strip()) tool_calls.append(call) except json.JSONDecodeError: print(f" [WARN] Could not parse tool call JSON: {match[:100]}") return tool_calls def run_agent(user_query: str, max_steps: int = 10) -> str: """ Transformers-native agentic loop — no vLLM or API server needed. Loop: 1. Generate response 2. Parse tool calls from output 3. Execute tools, append results 4. Repeat until no tool calls in response """ messages = [ { "role": "system", "content": ( "You are a helpful Singapore housing assistant. " "Always call the relevant tools to get accurate data before answering. " "Give a clear, concise summary after gathering all information." ), }, {"role": "user", "content": user_query}, ] print(f"\n{'='*60}") print(f"USER: {user_query}") print(f"{'='*60}") for step in range(max_steps): print(f"\n[Step {step + 1}] Generating...") response_text = generate_response(messages) print(f" Raw output: {response_text[:200]}...") # ── Try to parse tool calls from the response ── tool_calls = parse_tool_calls(response_text) if tool_calls: print(f" → Found {len(tool_calls)} tool call(s)") # ── Append assistant turn with raw response ── messages.append({ "role": "assistant", "content": response_text, }) # ── Execute each tool and append results ── for call in tool_calls: fn_name = call.get("name", "") fn_args = call.get("arguments", {}) # ── arguments may be a string or dict depending on model output ── if isinstance(fn_args, str): fn_args = json.loads(fn_args) print(f" • {fn_name}({json.dumps(fn_args, ensure_ascii=False)})") result = execute_tool(fn_name, fn_args) print(f" ↳ {result[:150]}") # ── Append tool result as tool role message ── messages.append({ "role": "tool", "name": fn_name, "content": result, }) continue # loop back for next generation # ── No tool calls — this is the final answer ── print(f"\n{'='*60}") print(f"AGENT FINAL ANSWER:\n{response_text}") print(f"{'='*60}\n") return response_text return "[Agent stopped: exceeded maximum steps]" # ── Run examples ── if __name__ == "__main__": run_agent( "Tolong carikan flat 4-bilik dekat Tampines, bajet bawah $500,000. " "Nak tahu juga berapa anggaran pinjaman bulanan." ) ``` Output ``` ============================================================ AGENT FINAL ANSWER: Tampines 4-room 500000 ============================================================ ``` ## Training Details ### Training Data 🤗[aisingapore/SEA-Instruct-2602](https://huggingface.co/datasets/aisingapore/SEA-Instruct-2602) ### Training Regime Our post-training workflow consists solely of distillation and model merging. ## Evaluation ### Testing Data, Factors & Metrics We evaluated Qwen-SEA-LION-v4.5 on general language, multi-turn chat and instruction-following capabilities. #### Testing Data **General language capabilities** For the evaluation of general language capabilities, we employed the SEA-HELM evaluation benchmark across a variety of tasks. These tasks include Question Answering (QA), Sentiment Analysis (Sentiment), Toxicity Detection (Toxicity), Translation in both directions (Eng>Lang & Lang>Eng), Abstractive Summarisation (Abssum), Causal Reasoning (Causal), Natural Language Inference (NLI), Linguistic Diagnostics (LINDSEA), Cultural Knowledge (Kalahi) and Global MMLU Lite. **Instruction-following and Multi-turn Chat** We evaluated the models on instruction-following and multi-turn chat capabilities with SEA-IFEval (based on IFEval) and SEA-MTBench (based on MT-Bench) respectively. The two datasets were originally in English, the linguists and native speakers in the team worked together to filter, localise and translate the datasets into the respective target languages to ensure that the examples remained reasonable, meaningful and natural. #### Factors All evaluations were run with the model specific generation parameters defined in the model config. Each evaluation comprised of 8 runs with different seeds and the final results were averaged across these runs. For all tasks, the model was expected to provide an answer tag from which the answer was automatically extracted. For tasks where options were provided, the answer should comprise one of the pre-defined options. The evaluation was done zero-shot with native prompts on a sample of 100-1000 instances for each dataset. - **SEA-IFEval:** SEA-IFEval evaluates a model's ability to adhere to constraints provided in the prompt, for example beginning a response with a specific word/phrase or answering with a certain number of sections. Additionally, accuracy is normalised by the proportion of responses in the correct language (if the model performs the task correctly but responds in the wrong language, it is judged to have failed the task). - **SEA-MTBench:** SEA-MTBench evaluates a model's ability to engage in multi-turn (2 turns) conversations and respond in ways that align with human needs. We use gpt-oss-120b as the judge model and compare against gpt-oss-120b as the baseline model. The metric used is the weighted win rate against the baseline model (i.e. average win rate across each category: Math, Reasoning, STEM, Humanities, Roleplay, Writing, Extraction). #### Metrics The following metrics were used for text capabilities: | **Task** | **Metric** | | --- | --- | | Sentiment Analysis | Accuracy | | Extractive QA (ID, VI, TH, TA) | ChrF++ | | MCQ-QA (TL, MY, MS) | Accuracy | | Metaphor | Accuracy | | Abstractive Summarisation | Rouge-L | | Translations | MetricX-24 score (with reference) | | Causal Reasoning | Accuracy | | Natural Language Inference | Accuracy | | LINDSEA | Accuracy | | Global MMLU Lite | Accuracy | | ThaiExam | Accuracy | | Kalahi | Accuracy | | SEA-IFEval | Accuracy | | SEA-MTBench | Win rate against a reference | #### Results ![LeaderboardResults!](Qwen-SEA-LION-v4.5-27B-IT_SEA-HELM_scores.png "LeaderboardCaptured20May") For details on Qwen-SEA-LION-v4.5-27B-IT performance, please refer to the [SEA-LION Leaderboard](https://leaderboard.sea-lion.ai/). *We are constantly updating the leaderboard - more to come very soon! #### Performance | GPU Chip | Model Size (GB) | VRAM Required (GB) | Time to First Token (s) | Tokens per Second | | --- | --- | --- | --- | --- | | H200 | 34.4 GB | 51.1 GiB | 0.0512 | 69.9005 | | H100 | 34.4 GB | 51.1 GiB | 0.326 | 49.03 | > *Additional Remarks:* > > - *TTFT and Tokens per Second: measured with vLLM on localhost and concurrency = 1.* > - *Offload all layers to GPU, Context Length 8192* > - *Reported results are the median (p50) values, calculated across 10 requests.* > - *Input size 4K, output 1K* > ## Technical Specifications ### Model Architecture The architecture is based on the highly efficient Qwen3.6 foundation. The detailed architecture can be found at . ## More Information This is the repository for the commercial instruction-tuned model. The model has not been aligned for safety. Developers and users should perform their own safety fine-tuning and related security measures. In no event shall the authors be held liable for any claims, damages, or other liabilities arising from the use of the released weights and codes. For more info, please contact us at [**sealion@aisingapore.org**](mailto:sealion@aisingapore.org) ## Team Ahmed Dabeer, Ahn Jeongmi, Antonyrex Sajeban, Chan Hok Teng Adwin, Cheng Zi Yi Nicholas, Choa Hsueh Mei Esther, Heng Jonathan, Huang Yuli, Jann Railey Estrada Montalan, Lee Chwan Ren, Leong Wai Yi, Leong Wei Qi, Liew Rachel, Limkonchotiwat Peerat, Muhammad Ridzuan Bin Mokhtar, Nagarajan Karthik, Ng Boon Cheong Raymond, Ngee Chia Tai, Ngui Jian Gang, Nguyen Thanh Ngan, Ong Tat-Wee David, Ong Zhi Hao, Pereira Mark, Poon Joseph, Rengarajan Hamsawardhini, Siow Wei Kang Bryan, Susanto Yosephine, Sutaveephamochanon Anocha, Tan Choon Meng, Tan Chor Phin Evelyn, Tan Siao Wei Jessica, Tan Yixian, Tee Jun Yun, Teng Kok Wai Walter, Teo Eng Sipp Leslie, Tjhi William, Wu Donghang, Yeo Yeow Tong, Yong Xianbin, Zhang Haoyang, Zhang Zhou ## Acknowledgement This project is supported by the National Research Foundation Singapore and Infocomm Media Development Authority (IMDA), Singapore under its National Large Language Model Funding Initiative. ## Contact [**sealion@aisingapore.org**](mailto:sealion@aisingapore.org)