DeepSeek V3 developed by the Chinese AI research lab DeepSeek under High-Flyer has been a standout in the AI landscape since its initial open-source release in December 2024. Known for its efficiency, performance, and accessibility, it continues to evolve rapidly. The latest update to DeepSeek V3, tagged \u201cDeepSeek V3 0324\u201d was rolled out on March 24, 2025, bringing subtle yet impactful refinements. Let\u2019s look at these updates and try the new DeepSeek V3 model. <\/p>\n
Minor Version Upgrade: DeepSeek V3 0324<\/h2>\n\n- The upgrade enhances the user experience across DeepSeek\u2019s official website, mobile app, and mini-program, with \u201cdeep thinking\u201d mode turned off by default. This suggests a focus on streamlining interaction rather than altering core capabilities.<\/li>\n
- The API interface and usage methods remain unchanged, ensuring continuity for developers. This means existing integrations (e.g., via model=\u2019deepseek-chat\u2019) don\u2019t require adjustments.<\/li>\n
- No major architectural changes were mentioned, indicating this is a refinement of the existing 671B-parameter Mixture-of-Experts (MoE) model, with 37B activated per token.<\/li>\n
- Availability: <\/strong>The updated model is live on the official DeepSeek platforms (website, app, mini-program) and HuggingFace.<\/a> The technical report and weights for \u201cDeepSeek V3 0324\u201d are accessible under the MIT license.<\/li>\n<\/ul>\n
How is DeepSeek V3 0324 Performing? <\/h2>\n
A user on X<\/a> tried the new DeepSeek V3 on my internal bench and it has a huge jump in all metrics on all tests. It is now the best non-reasoning model, dethroning Sonnet 3.5.<\/p>\n\n![\"\"](\"https:\/\/cdn.analyticsvidhya.com\/wp-content\/uploads\/2025\/03\/DeepSeek-V3-Performance.webp\")
Source: X<\/a><\/figcaption><\/figure>\n<\/div>\nDeepSeek V3 on Chatbot Arena leaderboard:<\/p>\n
How to Access the Latest DeepSeek V3? <\/h2>\n\n- Website: <\/strong>Test the updated V3 at deepseek.com<\/a> for free.<\/li>\n
- Mobile App:<\/strong> Available on iOS and Android, updated to reflect the March 24 release.<\/li>\n
- API: <\/strong>Use model=\u2019deepseek-chat\u2019 at api-docs.deepseek.com<\/a>. Pricing remains $0.14\/million input tokens (promotional until February 8, 2025, though an extension hasn\u2019t been ruled out).<\/li>\n
- HuggingFace: <\/strong>Download the \u201cDeepSeek V3 0324\u201d weights and technical report from here<\/a>. <\/li>\n<\/ul>\n
Let\u2019s Try the New DeepSeek V3 0324<\/h2>\n
I am going to use the updated DeepSeek model locally and via API.<\/p>\n
Using DeepSeek-V3-0324 Locally with llm-mlx Plugin<\/h3>\n
Installation Steps<\/strong><\/p>\nHere\u2019s what you need to run it on your machine (assuming you\u2019re using llm<\/code> CLI + mlx backend):<\/p>\n!pip install llm\n!llm install llm-mlx\n!llm mlx download-model mlx-community\/DeepSeek-V3-0324-4bit\n<\/code><\/pre>\nThis will:<\/p>\n
\n- Install the core
llm<\/code> CLI<\/li>\n- Add the MLX backend plugin<\/li>\n
- Download the 4-bit quantized model (
DeepSeek-V3-0324-4bit<\/code>) \u2014 more memory-efficient<\/li>\n<\/ol>\nRun a Chat Prompt Locally<\/strong><\/p>\nExample:<\/p>\n
!llm chat -m mlx-community\/DeepSeek-V3-0324-4bit 'Generate an SVG of a pelican riding a bicycle'<\/code><\/pre>\nOutput<\/strong>:<\/p>\n\n![\"\"](\"https:\/\/cdn.analyticsvidhya.com\/wp-content\/uploads\/2025\/03\/output.webp\")
<\/figure>\n<\/div>\nIf the model runs successfully, it should respond with an SVG snippet of a pelican on a bike \u2013 goofy and glorious.<\/p>\n
Using DeepSeek-V3-0324 via API<\/h3>\n
Install Required Package<\/strong><\/p>\n!pip3 install openai<\/code><\/pre>\nYes, even though you\u2019re using DeepSeek, you\u2019re interfacing with it using OpenAI-compatible SDK syntax.<\/p>\n
Python Script for API Interaction<\/strong><\/p>\nHere\u2019s a cleaned-up, annotated version of what\u2019s happening in the script:<\/p>\n
from openai import OpenAI\nimport time\n\n# Timing setup\nstart_time = time.time()\n\n# Initialize client with your DeepSeek API key and base URL\nclient = OpenAI(\n api_key=\"Your_api_key\", \n base_url=\"https:\/\/api.deepseek.com\" # This is important\n)\n\n# Send a streaming chat request\nresponse = client.chat.completions.create(\n model=\"deepseek-chat\",\n messages=[\n {\"role\": \"system\", \"content\": \"You are a helpful assistant\"},\n {\"role\": \"user\", \"content\": \"How many r's are there in Strawberry\"},\n ],\n stream=True\n)\n\n# Handle streamed response and collect metrics\nprompt_tokens = 0\ngenerated_tokens = 0\nfull_response = \"\"\n\nfor chunk in response:\n if hasattr(chunk, \"usage\") and hasattr(chunk.usage, \"prompt_tokens\"):\n prompt_tokens = chunk.usage.prompt_tokens\n\n if hasattr(chunk, \"choices\") and hasattr(chunk.choices[0], \"delta\") and hasattr(chunk.choices[0].delta, \"content\"):\n content = chunk.choices[0].delta.content\n if content:\n generated_tokens += 1\n full_response += content\n print(content, end=\"\", flush=True)\n\n# Performance tracking\nend_time = time.time()\ntotal_time = end_time - start_time\n\n# Token\/sec calculations\nprompt_tps = prompt_tokens \/ total_time if prompt_tokens > 0 else 0\ngeneration_tps = generated_tokens \/ total_time if generated_tokens > 0 else 0\n\n# Output metrics\nprint(\"\\n\\n--- Performance Metrics ---\")\nprint(f\"Prompt: {prompt_tokens} tokens, {prompt_tps:.3f} tokens-per-sec\")\nprint(f\"Generation: {generated_tokens} tokens, {generation_tps:.3f} tokens-per-sec\")\nprint(f\"Total time: {total_time:.2f} seconds\")\nprint(f\"Full response length: {len(full_response)} characters\")<\/code><\/pre>\nOutput<\/h3>\n### Final Answer**Answer:** There are **3 r's** in the word \"Strawberry.\"<\/p>
--- Performance Metrics ---
\n
How is DeepSeek V3 0324 Performing? <\/h2>\n
A user on X<\/a> tried the new DeepSeek V3 on my internal bench and it has a huge jump in all metrics on all tests. It is now the best non-reasoning model, dethroning Sonnet 3.5.<\/p>\n DeepSeek V3 on Chatbot Arena leaderboard:<\/p>\n I am going to use the updated DeepSeek model locally and via API.<\/p>\n Installation Steps<\/strong><\/p>\n Here\u2019s what you need to run it on your machine (assuming you\u2019re using This will:<\/p>\n Run a Chat Prompt Locally<\/strong><\/p>\n Example:<\/p>\n Output<\/strong>:<\/p>\n If the model runs successfully, it should respond with an SVG snippet of a pelican on a bike \u2013 goofy and glorious.<\/p>\n Install Required Package<\/strong><\/p>\n Yes, even though you\u2019re using DeepSeek, you\u2019re interfacing with it using OpenAI-compatible SDK syntax.<\/p>\n Python Script for API Interaction<\/strong><\/p>\n Here\u2019s a cleaned-up, annotated version of what\u2019s happening in the script:<\/p>\n **Answer:** There are **3 r's** in the word \"Strawberry.\"<\/p> --- Performance Metrics --- \nHow to Access the Latest DeepSeek V3? <\/h2>\n
\n
Let\u2019s Try the New DeepSeek V3 0324<\/h2>\n
Using DeepSeek-V3-0324 Locally with llm-mlx Plugin<\/h3>\n
llm<\/code> CLI + mlx backend):<\/p>\n!pip install llm\n!llm install llm-mlx\n!llm mlx download-model mlx-community\/DeepSeek-V3-0324-4bit\n<\/code><\/pre>\n\n
llm<\/code> CLI<\/li>\nDeepSeek-V3-0324-4bit<\/code>) \u2014 more memory-efficient<\/li>\n<\/ol>\n!llm chat -m mlx-community\/DeepSeek-V3-0324-4bit 'Generate an SVG of a pelican riding a bicycle'<\/code><\/pre>\n<\/figure>\n<\/div>\nUsing DeepSeek-V3-0324 via API<\/h3>\n
!pip3 install openai<\/code><\/pre>\nfrom openai import OpenAI\nimport time\n\n# Timing setup\nstart_time = time.time()\n\n# Initialize client with your DeepSeek API key and base URL\nclient = OpenAI(\n api_key=\"Your_api_key\", \n base_url=\"https:\/\/api.deepseek.com\" # This is important\n)\n\n# Send a streaming chat request\nresponse = client.chat.completions.create(\n model=\"deepseek-chat\",\n messages=[\n {\"role\": \"system\", \"content\": \"You are a helpful assistant\"},\n {\"role\": \"user\", \"content\": \"How many r's are there in Strawberry\"},\n ],\n stream=True\n)\n\n# Handle streamed response and collect metrics\nprompt_tokens = 0\ngenerated_tokens = 0\nfull_response = \"\"\n\nfor chunk in response:\n if hasattr(chunk, \"usage\") and hasattr(chunk.usage, \"prompt_tokens\"):\n prompt_tokens = chunk.usage.prompt_tokens\n\n if hasattr(chunk, \"choices\") and hasattr(chunk.choices[0], \"delta\") and hasattr(chunk.choices[0].delta, \"content\"):\n content = chunk.choices[0].delta.content\n if content:\n generated_tokens += 1\n full_response += content\n print(content, end=\"\", flush=True)\n\n# Performance tracking\nend_time = time.time()\ntotal_time = end_time - start_time\n\n# Token\/sec calculations\nprompt_tps = prompt_tokens \/ total_time if prompt_tokens > 0 else 0\ngeneration_tps = generated_tokens \/ total_time if generated_tokens > 0 else 0\n\n# Output metrics\nprint(\"\\n\\n--- Performance Metrics ---\")\nprint(f\"Prompt: {prompt_tokens} tokens, {prompt_tps:.3f} tokens-per-sec\")\nprint(f\"Generation: {generated_tokens} tokens, {generation_tps:.3f} tokens-per-sec\")\nprint(f\"Total time: {total_time:.2f} seconds\")\nprint(f\"Full response length: {len(full_response)} characters\")<\/code><\/pre>\nOutput<\/h3>\n
### Final Answer