DeepSeek-V4-Flash Update

Who Should Read

Backend developers weighing DeepSeek API or a model swap for a cheap coding-agent model

Core Mechanics

• Same model name, newer model
  • deepseek-flash now points to V4.1-Flash-1017, and the web and app models were updated too.
  • The API calling method and pricing are unchanged.
• Benchmarks rose over the previous V4.1
  • Terminal-Bench 4.0 went from 31.2 at the V4.1 Flash release to 42.3.
  • DeepSWE v1.1 went from 74.2 to 77.0, and ProgramBench from 20.3 to 29.1.
• Runtime and output length dropped
  • The post says tasks finish at lower cost and higher speed.
  • It claims a new performance-cost Pareto optimum across multiple harnesses (agent runtimes).
• Old model names route to V4.1 Flash
  • deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily routed to V4.1 Flash.
  • The previous-generation V4 Flash and Vision Exp models were retired.
• Thinking effort has three levels

Community Reactions

• Using Flash for 90% of tasks
  • A commenter keeps changes under 1000 lines and says they barely notice a difference from frontier models.
  • They use Flash for bug and security reviews too and cross-check the results.
• $4.55 for a month of usage
  • One commenter reports 3,467 requests and about 323M tokens over 30 days for personal agent work.
  • They are happy for coding and review and use other models for everything else.
• Mixing models in a workflow
  • One commenter implements with Flash and uses pricier models only for planning and review, about $0.5 for an hour of work.
  • Another prefers Codex or Claude for languages they do not know.
• Concerns about data and benchmarks
  • One comment worries that codebase information leaks to China via the DeepSeek API.
  • Another asks whether these models are just fine-tuned for benchmarks.

How to Apply

• Swap the model name to deepseek-flash and compare
  • Change your deepseek-v4-flash calls to deepseek-flash and run the same 10 tasks today.
  • Record correctness, time, and output tokens next to the earlier results.
• Split thinking effort by task difficulty
  • Use low for simple tasks like classification or summaries and high for daily agent work.
  • Raise only the frequently failing complex tasks to max and log the cost difference.
• Plan with a big model and implement with Flash
  • Use a pricier model for planning and review, and Flash for implementation and log analysis.
  • Measure cost per stage to confirm the split is actually cheaper.
• Check data policy before sending private code
  • Confirm with your security team whether internal code may go to an external API.
  • If not, evaluate with public code or sample data only.

Try It

# Paste into Claude or ChatGPT (a simplified evaluation prompt based on the post)

You are an LLM API cost and quality evaluation designer. Our service is {{service description}}, and our current model is {{current model}}.
We want to decide whether to switch to DeepSeek's deepseek-flash.

Create the following.
1. Two test cases for each of 5 task types in our service
2. A one-line pass criterion for each case
3. A table to record when running with thinking effort low/high/max (correct or not, time, output tokens, cost)

Task types: {{task type list}}

Next step: paste the filled table back and ask for the cheapest effort setting per task type.

Terminology

Related Papers

• Mercury 3: Diffusion at scale

  Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality

• Timestral: Mistral's time-series foundation models

  A 209M time-series model ranks first on three benchmarks and runs on CPU.

• DeepSeek V4.1 Flash 1017 (Max)

  DeepSeek V4.1 Flash 1017 Max scores 45 on the Intelligence Index at 211 tok/s and $0.30 input.

• If you don't run your own LLM proxy, how do you know whats going on?

  Running your own LLM proxy shows which keys, models, and cache misses drive your usage.

• DeepSeek: DeepSeek v4.1 Flash 1017

  DeepSeek V4.1 Flash 1017: $0.05/M input, 1M context, built for agents

• DeepSeek-v4.1-Flash-1017

  DeepSeek V4.1 Flash minor update lifts coding and agent benchmarks while cutting run time and output length

• Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware

Related Resources

• https://api-docs.deepseek.com/updates/
• https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731