Skip to content
BSWEN
Blog

DeepSeek V4 Flash vs MiniMax for Coding Agents: Which LLM Follows Instructions Better?

Cowrie
Cowrie
Dev @ Bswen

Problem

When building a coding agent that handles multi-file edits, shell commands, and frequent back-and-forth interactions, the choice of LLM determines whether the agent actually works or just wastes tokens. I need a model that follows instructions reliably — not one that calls tools endlessly without purpose.

I tested MiniMax (M2.5 and M2.7 variants) first. With an 80-line system prompt — less than a quarter of what I planned to use — it struggled. It broke mid-execution, called tools in “action without purpose” patterns, and required constant manual fixes. The agent was more work than doing things manually.

The Comparison

Here are the real-world metrics from a 3-week coding agent workload:

MetricDeepSeek V4 FlashMiniMax (M2.5/M2.7)
System prompt length330 lines80 lines
Total API calls13,97813,389
Tokens processed1.7B794M
Total cost$10.37$52.87
Cache hit rate98%~70%
ReliabilityPerfect adherenceBroke mid-execution

DeepSeek V4 Flash handled 4x the system prompt complexity without losing context or deviating from instructions. And it cost 5x less.

What Happened?

I set up the coding agent with a comprehensive system prompt covering tool usage patterns, file editing conventions, shell command safety, and error handling. The prompt was 330 lines — detailed but not unusual for agentic workflows.

Before settling on DeepSeek, I tried MiniMax first. I kept the system prompt shorter (80 lines) because MiniMax’s documentation suggested simpler prompts work better. But even with minimal instructions, the agent behaved erratically:

MiniMax behavior patterns
- Called read-file on the same file 5 times in 3 seconds
- Generated code, then immediately called delete-file on it
- Reached the same logical conclusion 3 times via different tool paths
- Broke the edit sequence mid-operation and started writing random files

Every session required supervision. I spent more time fixing the agent’s mistakes than the agent saved me.

I switched to DeepSeek V4 Flash. I increased the system prompt to 330 lines — including everything MiniMax couldn’t handle and more. The result: reliable, consistent behavior across 13,978 calls. No mid-execution breaks, no endless tool loops, no random file writes.

Why DeepSeek Works Better for Agents

I think the key reason is that DeepSeek’s training and architecture prioritize prompt adherence. The model treats the system prompt as a binding contract — not as a suggestion.

For coding agents, this matters more than any benchmark score:

  • Tool calling discipline: DeepSeek calls tools only when needed, not as a nervous habit
  • Context retention: 330 lines of instructions with no degradation over 13,978 calls
  • Error recovery: When something goes wrong, DeepSeek follows the recovery instructions in the prompt instead of inventing its own (usually wrong) fix
  • Cost discipline: Better instruction following means fewer wasted calls — which is why DeepSeek costs less despite processing more tokens

Common Mistakes

Assuming all LLMs are interchangeable for agentic workloads is the biggest mistake. MiniMax appeared competitive in standalone benchmarks but failed in practice. Always test agent-specific workloads rather than relying on general NLP benchmarks.

Another mistake: optimizing for per-token price without considering instruction-following quality. A cheaper model that produces bad output wastes money on retries and developer time.

Summary

In this post, I compared DeepSeek V4 Flash and MiniMax for coding agent workloads based on a real-world 3-week test. The key takeaway is that DeepSeek V4 Flash offers superior instruction following at 5x lower cost. For agentic workflows in mid-2026, DeepSeek V4 Flash is the practical choice.

Final Words + More Resources

My intention with this article was to help others share my knowledge and experience. If you want to contact me, you can contact by email: Email me

Here are also the most important links from this article along with some further resources that will help you in this scope:

Oh, and if you found these resources useful, don’t forget to support me by starring the repo on GitHub!

Comments

Tags: