Update README.md

README.md

@@ -1,43 +1,58 @@
----
-license: apache-2.0
----
-# Reka Flash 3.1
-At Reka, we build intelligence from the ground up to power our multimodal platform. From atomic capabilities such as [visual understanding](https://www.reka.ai/news/reka-visual-language-models) and [reasoning to use tools](https://www.reka.ai/news/introducing-reka-flash) to system-level optimization to serve them at scale.
+Reka Flash 3.1 is a 21B general-purpose reasoning model that was trained from scratch. It was trained in synthetic and public datasets for supervised finetuning, followed by large-scale RLOO with rule-based rewards. Reka Flash 3.1 is an improved version of Reka Flash 3 due to significant advances in our reinforcement learning stack and curated high-qaulity RL data. Reka Flash 3.1 is particularly strong on coding and as a base model to be finetuned on agentic tasks.  
+Reka Flash 3.1 improves by 10 points on LiveCodeBench v5 (Full set) from Reka Flash 3. For coding related tasks, Reka Flash 3.1 is competitive with models such as Qwen3-32B. o3-mini, and Gemini 2.5 Flash Thinking. If you want to learn more about how we do reinforcement learning for Reka Flash 3.1 that results in these improvements, please check out this post.
 
-Today, we are excited to open source a few of our building blocks:  
-**Reka Flash 3.1**, an improved version of Reka Flash 3 due to significant advances in our reinforcement learning stack. Reka Flash 3.1 is particularly strong on coding and as a base model to be finetuned on agentic tasks.  
-**Reka Flash Quantized (Reka Quant)**, a 3.5-bit quantized version of Reka Flash 3.1 that delivers state-of-the-art performance at low bitwidths using calibrated error reduction and self distillation.  
-Our quantization library supports self-distillation, fast distributed proxy Hessian computation for fast LDLQ, and export to popular llama.cpp datatypes such as Q3_K and Q4_K.  
+![image/png](https://cdn-uploads.huggingface.co/production/uploads/65a776598ee30c06716aa380/e5SsUG4vTgslFtcJxM9DT.png)
 
+Try it out at [Reka Space](https://space.reka.ai).
 
+Strong reasoning and coding skills are important capabilities to support multimodal agentic use cases, and near-lossless quantization allows us to deploy our models anywhere. A multimodal version of Reka-Flash-3.1 serves as a base model for our core products Reka Research and Reka Vision. Please contact us for more information about how you can use them in your organizations.
 
-Reka Flash 3.1 improves by 10 points on LiveCodeBench v5 (Full set) from Reka Flash 3. For coding related tasks, Reka Flash 3.1 is competitive with models such as Qwen3-32B. o3-mini, and Gemini 2.5 Flash Thinking. If you want to learn more about how we do reinforcement learning for Reka Flash 3.1 that results in these improvements, please check out this post.  
+Model efficiency is critical for the local deployment. We also release a quantized version of Reka Flash 3.1 in this link. Meanwhile, we opensource the corresponding quantizatioon library at this link.
 
-While Reka Flash 3.1 is already compact as a 21 billion parameter model, quantization allows us to reduce its memory footprint even further, allowing it to work in resource-constrainted settings and be served cost efficiently. Reka Quant achieves near-lossless quantization to 3.5 bits when quantizing Reka Flash 3.1 to Q3_K_S datatype in llama.cpp, incurring only a 1.6 average performance degradation. In contrast, Q3_K_S quantization routine results in a 6.8 average performance degradation. We provide a more detailed discussion about our quantization approach in this post.  
 
-Strong reasoning and coding skills are important capabilities to support multimodal agentic use cases, and near-lossless quantization allows us to deploy our models anywhere. A multimodal version of Reka-Flash-3.1 serves as a base model for our core products Reka Research and Reka Vision. Please contact us for more information about how you can use them in your organizations.
 
-![image/png](https://cdn-uploads.huggingface.co/production/uploads/65a776598ee30c06716aa380/e5SsUG4vTgslFtcJxM9DT.png)
 
-## Quick Start
-For ease of deployment, Reka Flash 3.1 is released in a Llama-compatible format. You may use any library compatible with Llama to run the model.  
+## Quickstart
 
-### Via Huggingface
-```python
-import transformers
-model_name = "RekaAI/reka-flash-3.1"
+For ease of deployment, the model is released in a Llama-compatible format. You may use any library compatible with Llama to run the model.
 
-tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)
-model = transformers.AutoModelForCausalLM.from_pretrained(model_name, torch_dtype='auto', device_map='auto')
+### Via Hugging Face
 
+```python
+import transformers
+tokenizer = transformers.AutoTokenizer.from_pretrained("RekaAI/reka-flash-3.1")
+model = transformers.AutoModelForCausalLM.from_pretrained("RekaAI/reka-flash-3.1", torch_dtype='auto', device_map='auto')
 prompt = {"role": "human", "content": "Write a poem about large language model."}
 text = tokenizer.apply_chat_template([prompt], tokenize=False, add_generation_prompt=True)
 model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
-outputs = model.generate(**model_inputs, max_new_tokens=65536)
+outputs = model.generate(**model_inputs, max_new_tokens=512)
 print(tokenizer.decode(outputs[0], skip_special_tokens=True))
 ```
 
 ### Via vLLM
-```shell
+
+```bash
 docker run --rm -it --network=host --gpus '"device=0"' -v  --shm-size=10.24gb vllm/vllm-openai:latest serve RekaAI/reka-flash-3.1 --dtype auto -tp 1
 ```
+
+## Model Details
+
+### Prompt Format
+
+Reka Flash 3.1 uses cl100k_base tokenizer and adds no additional special tokens. Its prompt format is as follows:
+```
+human: this is round 1 prompt <sep> assistant: this is round 1 response <sep> ...
+```
+Generation should stop on seeing the string `<sep>` or seeing the special token `<|endoftext|>`.
+System prompt can be added by prepending to the first user round.
+```
+human: You are a friendly assistant blah ... this is round 1 user prompt <sep> assistant: this is round 1 response <sep> ...
+```
+For multi-round conversations, it is recommended to drop the reasoning traces in the previous assistant round to save tokens for the model to think.
+If you are using HF or vLLM, the built-in chat_template shall handle prompt formatting automatically.
+
+
+
+### Language Support
+
+This model is primarily built for the English language, and you should consider this an English only model. However, the model is able to converse and understand other languages to some degree.