MiniCPM5-2B vs Qwen3.5 4B
Wins 2 of 6 areas
Coding · Following instructions
Wins 2 of 6 areas
Reasoning · Long documents
The two are evenly matched.MiniCPM5-2B is better at coding and following instructions; Qwen3.5 4B at reasoning and long documents.
Scores updated · 30 tests both models report · How we compare
Where each one wins
Tests won in each of the six areas where both have results. Each piece is one test, so a longer bar means more evidence; grey means the two scored within a point of each other.
- ReasoningHard problems that need careful thinking02Qwen3.5 4B2 of 3 tests · 1 tie
- Long documentsFinding answers in very long texts01Qwen3.5 4B1 of 1 test
- CodingWriting and fixing software21MiniCPM5-2B2 of 3 tests
- Following instructionsDoing exactly what it is asked10MiniCPM5-2B1 of 1 test
- AgentsCarrying out multi-step tasks on its own11Even1 each
- FactsGetting facts right instead of making them up11Even1 each
Long documents and following instructions rest on a single test each.
What it costs
Prices per million tokens, roughly 750,000 words. The bars show the cost of a million tokens read plus a million written.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where Qwen3.5 4B pulls ahead
- Command-line tasks in a real terminalTerminal-Bench 2.1+17.2points ahead
- Multi-step tasks using many real software toolsMCP Atlas+15points ahead
- Graduate-level biology, physics and chemistry questionsGPQA Diamond+6.9points ahead
Where MiniCPM5-2B pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+64.7points ahead
- Recent programming contest problemsLiveCodeBench v6+14.7points ahead
- Customer-service tasks in a simulated bankτ-Bench V3 · Banking+14points ahead
Every test, side by side
All 30 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingMiniCPM5-2B
- Terminal-Bench 2.1Qwen3.5 4B by 17.28.625.8+17.2
- LiveCodeBench v6MiniCPM5-2B by 14.770.555.8+14.7
- SciCodeMiniCPM5-2B by 10.226.316.1+10.2
AgentsEven
- MCP AtlasQwen3.5 4B by 1526.441.4+15
- τ-Bench V3 · BankingMiniCPM5-2B by 1420.86.8+14
ReasoningQwen3.5 4B
- GPQA DiamondQwen3.5 4B by 6.970.277.1+6.9
- Humanity's Last ExamQwen3.5 4B by 18.99.9+1
- CritPttie0.30tie
FactsEven
- AA-Omniscience · Non-hallucinationMiniCPM5-2B by 64.778.113.4+64.7
- AA-Omniscience · AccuracyQwen3.5 4B by 6.78.415.1+6.7
Following instructionsMiniCPM5-2B
- IFBenchMiniCPM5-2B by 10.162.152+10.1
Other results18 tests, not counted
Tests outside the eight areas. They are not counted above: several are summary scores built from other tests, or the same test under another name.
- AA-OmniscienceMiniCPM5-2B by 46.7-11.7-58.4+46.7
- RULER 128KQwen3.5 4B by 46.63884.6+46.6
- AA Agentic IndexQwen3.5 4B by 23.5932.5+23.5
- WebArena-VerifiedQwen3.5 4B by 12.89.121.9+12.8
- EnterpriseOps-GymQwen3.5 4B by 12.613.426+12.6
- BFCLv4MiniCPM5-2B by 12.162.650.6+12.1
- Claw-Eval General (EN), pass^3Qwen3.5 4B by 8.434.643+8.4
- Artificial Analysis Coding IndexQwen3.5 4B by 8.114.522.6+8.1
- BrowseComp-Plus (OpenClaw subset)Qwen3.5 4B by 7.916.824.7+7.9
- Terminal-Bench 2.1 (OpenCode)Qwen3.5 4B by 7.95.613.5+7.9
- SWE-bench Multilingual (OpenCode)Qwen3.5 4B by 7.615.723.3+7.6
- LiveCodeBench Pro 25Q2MiniCPM5-2B by 529.324.3+5
- AutomationBench PublicQwen3.5 4B by 44.78.7+4
- τ³-bench (4 domains), averageMiniCPM5-2B by 3.469.466+3.4
- IFEvalQwen3.5 4B by 287.889.8+2
- Claw-Eval General (EN), averageQwen3.5 4B by 1.360.762+1.3
- AA Intelligencetie12.513.1tie
- HLE (with tools)tie11.511.6tie
Questions people ask
Which is better, MiniCPM5-2B or Qwen3.5 4B?
MiniCPM5-2B and Qwen3.5 4B each win two of the six areas where both have results. MiniCPM5-2B wins coding and following instructions; Qwen3.5 4B wins reasoning and long documents. They are level on agents and facts.
Which is better for coding?
MiniCPM5-2B. It wins 2 of the 3 coding tests both models report; Qwen3.5 4B wins 1.
How do you compare the two?
We use the 30 benchmark tests both models have published scores on. The verdict counts the 12 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 18 are listed but not counted, because several are summary scores or repeat a test. Each score is the one shown on the model's own page.