Back to the lab
Benchmark

Beyond the Leaderboard: DeepSeek-V4-Pro

A thinking model that negotiates constraints and writes clean code — but carries an 8× speed penalty versus Kimi.

DeepSeekbenchmarkBeyond the LeaderboardOllama Cloudreasoning

Beyond the Leaderboard: DeepSeek-V4-Pro

Model: ollama/deepseek-v4-pro:cloud
Test date: 2026-06-02
Full write-up: SMF Works blog

What we tested

DeepSeek-V4-Pro is the model SMF Works reaches for when a task needs careful reasoning. As a "thinking" model it emits an internal chain-of-thought before answering. We ran the same 15-test production harness to see whether the extra time translated into better results.

Results at a glance

Metric Value
Overall score 0.72
Tests passed 6/15
Avg time-to-first-token ~17.5s
Avg total time ~35s
Reliability 100%

Test-by-test results

Test Score Passed Key finding
Basic Reasoning 0.70 Correct, verbose
Code Generation 0.70 Clean, typed, good edge handling
Debugging 0.50 Correctly said code was fine
Algorithm Explanation 0.35 2 long sentences, missed 3-sentence target
Complex Multi-Step Reasoning 0.25 Wrong answer, hit token limit
Content Generation 0.50 Drifted into operational context
Edge Case Handling 0.50 Asked clarifying questions
Long-Context RAG 0.50 Got McKinsey stat, missed attribution
Structured Output (JSON) 1.00 Perfect JSON
Tool Use 0.50 Listed calls, no execution
Instruction Following 0.70 2/5 constraints met
Adversarial / Trick 0.75 Correct (5 minutes)
Code Execution Reasoning 0.88 Correct outputs
Summarization Fidelity 0.50 Missed key facts
Recent Knowledge 0.50 Accurate May 2025 cutoff

What worked

  • Instruction following is the standout win. DeepSeek scored 0.70 versus Kimi's 0.50 by visibly negotiating constraints (substituting words to reduce "e" usage).
  • Code generation is production-ready, slightly cleaner than Kimi.
  • JSON output is perfect.
  • Calibration on knowledge cutoff is honest — it correctly stated May 2025.

What didn't

  • Speed. 17.5s TTF is 8× slower than Kimi. Conversational or real-time use is painful.
  • Complex reasoning still hit the token limit without converging.
  • Algorithm explanation was worse than Kimi because it produced dense, run-on sentences.

Verdict

Use DeepSeek-V4-Pro when precision matters more than speed: research synthesis, careful code review, or prompts with multiple simultaneous constraints. Do not use it for real-time interfaces or cron jobs that collide.