DeepSeek V4.1 Flash vs Step 5 Preview
DeepSeek V4.1 Flash and Step 5 Preview match in llmrun Score, with values of 129.4 versus 128.0; neither fits 24 GB.
Model A · DeepSeek
DeepSeek V4.1 Flash763.2B · 16B active · 1M context · MIT · Sep 2026
Model B · StepFun
Step 5 Preview604.3B · 27B active · 1M context · Other · Sep 2026
Choose DeepSeek V4.1 Flash if you want…
Also does vision
Step 5 Preview doesn't list it
More permissive license
MIT vs Other
Choose Step 5 Preview if you want…
Stronger at science
134.2 vs 122.7
Also does chat
DeepSeek V4.1 Flash doesn't list it
On your hardware
Each model at its default quant (Q4_K_M), a smaller one when it must. Estimates, ±20%.
8 GB
RTX 4060 · 8 GB tier
DeepSeek V4.1 Flash
Doesn't fit
smallest quant needs 210.4 GB
Step 5 Preview
Doesn't fit
smallest quant needs 191.4 GB
12 GB
RTX 3060 12GB · 12 GB tier
DeepSeek V4.1 Flash
Doesn't fit
smallest quant needs 210.4 GB
Step 5 Preview
Doesn't fit
smallest quant needs 191.4 GB
16 GB
RTX 5060 Ti 16GB · 16 GB tier
DeepSeek V4.1 Flash
Doesn't fit
smallest quant needs 210.4 GB
Step 5 Preview
Doesn't fit
smallest quant needs 191.4 GB
24 GBreference
RTX 4090 · 24 GB tier
DeepSeek V4.1 Flash
Doesn't fit
smallest quant needs 210.4 GB
Step 5 Preview
Doesn't fit
smallest quant needs 191.4 GB
32 GB
RTX 5090 · 32 GB tier
DeepSeek V4.1 Flash
Doesn't fit
smallest quant needs 210.4 GB
Step 5 Preview
Doesn't fit
smallest quant needs 191.4 GB
48 GB
RTX A6000 · 48 GB tier
DeepSeek V4.1 Flash
Doesn't fit
smallest quant needs 210.4 GB
Step 5 Preview
Doesn't fit
smallest quant needs 191.4 GB
Mac 32 GB
Mac Mini M4 · Macs
DeepSeek V4.1 Flash
Doesn't fit
smallest quant needs 210.4 GB
Step 5 Preview
Doesn't fit
smallest quant needs 191.4 GB
Mac 64 GB
Mac Studio M4 Max · Macs
DeepSeek V4.1 Flash
Doesn't fit
smallest quant needs 210.4 GB
Step 5 Preview
Doesn't fit
smallest quant needs 191.4 GB
Mac 128 GB
Mac Studio M4 Max · Macs
DeepSeek V4.1 Flash
Doesn't fit
smallest quant needs 210.4 GB
Step 5 Preview
Doesn't fit
smallest quant needs 191.4 GB
Reference GPU: RTX 4090. Tinted cell = the better pick on that machine (fits, more headroom, or at least 10% faster).
Overall
DeepSeek V4.1 Flash
Step 5 Preview
Coding
DeepSeek V4.1 Flash
Step 5 Preview
no data
Agents
DeepSeek V4.1 Flash
Step 5 Preview
no data
Math
DeepSeek V4.1 Flash
Step 5 Preview
no data
Science
DeepSeek V4.1 Flash
Step 5 Preview
Reasoning
DeepSeek V4.1 Flash
Step 5 Preview
no data
DeepSeek V4.1 Flash · #6 of 101 open models · #5 of 98 that fit ≤512 GB
Step 5 Preview · #8 of 101 open models · #6 of 98 that fit ≤512 GB
Benchmarks
3 shared boards: DeepSeek V4.1 Flash leads 1, Step 5 Preview leads 2 · Benchmarks imported Oct 2026
DeepSeek V4.1 Flash
51.8%#7/68 open
Lab-reportedEpoch AI · Sep 2026
Step 5 Preview
58.9%#4/68 open
Lab-reportedEpoch AI · Sep 2026
DeepSeek V4.1 Flash
14.3%#10/68 open
Lab-reportedEpoch AI · Sep 2026
Step 5 Preview
20.9%#3/68 open
Lab-reportedEpoch AI · Sep 2026
DeepSeek V4.1 Flash
54.0%#6/27 open
Lab-reportedEpoch AI · Sep 2026
Step 5 Preview
42.0%#9/27 open
Lab-reportedEpoch AI · Sep 2026
30 more boards (other shared and single-model boards)
Only DeepSeek V4.1 Flash (30)
- GPQA Diamond89.8%
- LMArena Text1476.0
- LiveBench Coding80.0
- LiveBench Reasoning86.7
- AIME 2024/202598.3%
- FrontierMath67.4%
- LMArena WebDev1618.6
- LiveBench Math93.3
- ARC-AGI-272.9%
- ARC-AGI94.5%
- SimpleQA42.9%
- FrontierMath Tier 426.8%
- FrontierCode43.9%
- APEX-Agents39.5%
- Chess Puzzles48.0%
- DTBench89.9%
- Furniture Assembly34.2%
- LiveBench 2026 Reasoning86.7
- LMCA47.0%
- Mystery Game Puzzles43.0%
- GDP.pdf19.8%
- LiveBench 2026 Browser84.0
- LiveBench 2026 Data Analysis79.3
- ALE-Bench1092.3
- LiveBench 2026 Coding90.2
- LiveBench 2026 Long-Horizon40.1
- Surface Evolver Bench46.3%
- LiveBench 2026 Design79.2
- WeirdML v35.9%
- LiveBench 2026 Research86.8
VRAM by quantization
Memory to load the model and 2K context; below it, the total at 32K and 128K context. Lighter side in bold.
Q2_K
DeepSeek V4.1 Flash
325 GB264.5 GB file
32K 325 GB · 128K 326 GB
Step 5 Preview
283 GB182.5 GB file
32K 301 GB · 128K 360 GB
Q3_K_M
DeepSeek V4.1 Flash
372 GB347.3 GB file
32K 373 GB · 128K 374 GB
Step 5 Preview
324 GB~294.6 GB file
32K 343 GB · 128K 402 GB
Q4_K_M
DeepSeek V4.1 Flash
458 GB444.7 GB file
32K 459 GB · 128K 460 GB
Step 5 Preview
399 GB~362.6 GB file
32K 417 GB · 128K 476 GB
Q5_K_M
DeepSeek V4.1 Flash
544 GB~543.8 GB file
32K 545 GB · 128K 546 GB
Step 5 Preview
474 GB~430.6 GB file
32K 492 GB · 128K 551 GB
Q6_K
DeepSeek V4.1 Flash
630 GB~629.6 GB file
32K 630 GB · 128K 632 GB
Step 5 Preview
548 GB~498.6 GB file
32K 567 GB · 128K 626 GB
Q8_0
DeepSeek V4.1 Flash
764 GB508.0 GB file
32K 764 GB · 128K 765 GB
Step 5 Preview
665 GB~604.3 GB file
32K 683 GB · 128K 742 GB
4 more quants
IQ2_XXS
DeepSeek V4.1 Flash
210 GB~209.9 GB file
32K 211 GB · 128K 212 GB
Step 5 Preview
183 GB~166.2 GB file
32K 201 GB · 128K 260 GB
IQ3_XS
DeepSeek V4.1 Flash
315 GB434.0 GB file
32K 316 GB · 128K 317 GB
Step 5 Preview
not listed
IQ3_M
DeepSeek V4.1 Flash
not listed
Step 5 Preview
299 GB271.6 GB file
32K 318 GB · 128K 377 GB
BF16
DeepSeek V4.1 Flash
1527 GB~1526.4 GB file
32K 1527 GB · 128K 1528 GB
Step 5 Preview
1330 GB~1208.7 GB file
32K 1348 GB · 128K 1407 GB
DeepSeek V4.1 Flash
Ollama · Q4_K_M
ollama run hf.co/vcruz305/DeepSeek-V4.1-Flash-GGUF:Q4_K_MStep 5 Preview
Ollama · Q4_K_M
ollama run hf.co/smalinin/Step-5-Preview-GGUF| Spec | DeepSeek V4.1 Flash | Step 5 Preview |
|---|---|---|
| Publisher | DeepSeek | StepFun |
| Parameters | 763.2B | 604.3B |
| Active parameters | 16B per token | 27B per token |
| Architecture | DeepseekV41ForCausalLM | Step5ForConditionalGeneration |
| Context | 1M tokens | 1M tokens |
| Modalities | Vision, Function-calling | Chat, Function-calling |
| License | MIT | Other |
| Released | Sep 2026 | Sep 2026 |
Frequently Asked Questions
Which is better, DeepSeek V4.1 Flash or Step 5 Preview?
DeepSeek V4.1 Flash and Step 5 Preview match in llmrun Score, with values of 129.4 versus 128.0; neither fits 24 GB. Choose DeepSeek V4.1 Flash for also does vision and more permissive license; choose Step 5 Preview for stronger at science and also does chat.
Which needs less VRAM, DeepSeek V4.1 Flash or Step 5 Preview?
At Q4_K_M, DeepSeek V4.1 Flash requires 458.3 GB while Step 5 Preview needs 398.9 GB, making Step 5 Preview the easier model to run on your local setup.
Can I run DeepSeek V4.1 Flash or Step 5 Preview on a 16 GB GPU?
DeepSeek V4.1 Flash does not fit (smallest quant needs 210.4 GB); Step 5 Preview does not fit (smallest quant needs 191.4 GB).
Which has a longer context window, DeepSeek V4.1 Flash or Step 5 Preview?
DeepSeek V4.1 Flash supports 1,048,576 tokens and Step 5 Preview supports 1,048,576 tokens.