anocha-mcomp commited on
Commit
eb2faab
·
verified ·
1 Parent(s): bcec397

Update model card from Confluence: Qwen-SEA-LION-v4.5-27B Draft

Browse files
Files changed (1) hide show
  1. qwen-sea-lion-v45-27b_draft.md +15 -13
qwen-sea-lion-v45-27b_draft.md CHANGED
@@ -1,20 +1,22 @@
 
 
 
 
1
  # Model Card for Qwen-SEA-LION-v4.5-27B
2
 
3
  Last update: 2026-05-19
4
 
5
  **SEA-LION** is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned for the Southeast Asia (SEA) region.
6
 
7
- **Qwen-SEA-LION-v4.5-27B** 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 Qwen/Qwen3.5-397B-A17B on an updated aisingapore/SEA-Instruct-2602, instilling multilingual and multicultural fluency across English and key SEA languages: Burmese, Indonesian, Filipino, Malay, Tamil, Thai, and Vietnamese.
8
 
9
  Qwen-SEA-LION-v4.5-27B inherits the following features from Qwen3.6:
10
 
11
- - **Context Window (262K):** Reduce if facing OOM errors but keep ≥128K to preserve full reasoning capabilities.
12
- - **Unified Vision-Language:** Early fusion training delivers good performance across multimodal reasoning, coding, and visual tasks.
13
- - **Scalable RL:** Trained in million-agent environments for robust, real-world SEA adaptability.
14
- - **Broad Linguistic Coverage:** Deeply specialized in SEA cultural nuances while supporting 201 languages globally.
15
- - **Advanced Infrastructure:** Utilizes highly efficient multimodal training and asynchronous RL frameworks.
16
- - **Agentic Coding:** High-precision handling of repository-level reasoning and frontend workflows.
17
  - **Thinking Preservation:** Retains historical reasoning context to streamline iterative development and reduce compute overhead.
 
 
18
 
19
  ## Model Details
20
 
@@ -22,7 +24,7 @@ Qwen-SEA-LION-v4.5-27B inherits the following features from Qwen3.6:
22
 
23
  SEA-LION stands for Southeast Asian Languages In One Network.
24
 
25
- 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.
26
 
27
  For tokenization, the model employs the default tokenizer used in Qwen3.6.
28
 
@@ -557,12 +559,12 @@ For details on Qwen-SEA-LION-v4.5-27B performance, please refer to the [SEA-LION
557
  | H200 | 34.4 GB | 51.1 GiB | 0.0512 | 69.9005 |
558
  | H100 | 34.4 GB | 51.1 GiB | | |
559
 
560
- > Additional Remarks:
561
  >
562
- > - TTFT and Tokens per Second: measured with vllm on localhost and concurrency = 1.
563
- > - Offload all layers to GPU, Context Length 8192
564
- > - Reported results are the median (p50) values, calculated across 10 requests.
565
- > - Input size 4K, output 1K
566
  >
567
 
568
  ## Technical Specifications
 
1
+ # Qwen-SEA-LION-v4.5-27B Draft
2
+
3
+ `![Banner!](Qwen-SEA-LION-v4.5.png "Qwen-SEA-LION-v4.5-27B")`
4
+
5
  # Model Card for Qwen-SEA-LION-v4.5-27B
6
 
7
  Last update: 2026-05-19
8
 
9
  **SEA-LION** is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned for the Southeast Asia (SEA) region.
10
 
11
+ **Qwen-SEA-LION-v4.5-27B** 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.
12
 
13
  Qwen-SEA-LION-v4.5-27B inherits the following features from Qwen3.6:
14
 
15
+ - **Context Window (262K):** Large context window to enable strong reasoning capabilities.
16
+ - **Reasoning** – Highly capable reasoning model, with configurable thinking modes.
 
 
 
 
17
  - **Thinking Preservation:** Retains historical reasoning context to streamline iterative development and reduce compute overhead.
18
+ - **Agentic Coding:** High-precision handling of repository-level reasoning and frontend workflows.
19
+ - **Unified Vision-Language:** Early fusion training delivers good performance across multimodal reasoning, coding, and visual tasks.
20
 
21
  ## Model Details
22
 
 
24
 
25
  SEA-LION stands for Southeast Asian Languages In One Network.
26
 
27
+ 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.
28
 
29
  For tokenization, the model employs the default tokenizer used in Qwen3.6.
30
 
 
559
  | H200 | 34.4 GB | 51.1 GiB | 0.0512 | 69.9005 |
560
  | H100 | 34.4 GB | 51.1 GiB | | |
561
 
562
+ > *Additional Remarks:*
563
  >
564
+ > - *TTFT and Tokens per Second: measured with vllm on localhost and concurrency = 1.*
565
+ > - *Offload all layers to GPU, Context Length 8192*
566
+ > - *Reported results are the median (p50) values, calculated across 10 requests.*
567
+ > - *Input size 4K, output 1K*
568
  >
569
 
570
  ## Technical Specifications