Head to head

Gemma 4 31B vs DeepSeek V4.1 Flash

DeepSeek V4.1 Flash scores higher on 3 of the 3 benchmarks both models publish. The widest gap is on Humanity's Last Exam (with tools), where DeepSeek V4.1 Flash leads by 37.4 points. Both have open weights; Gemma 4 31B is far lighter to host, needing about 21 GB at 4-bit against 381 GB for DeepSeek V4.1 Flash.

Gemma 4 31BGoogle · OpenLLM Buddy catalog79.2compositeDeepSeek V4.1 FlashDeepSeek · OpenLLM Buddy catalog79.8composite

Shared benchmarks

Raw published scores; the higher score on each row is highlighted. LMArena Elo is a human-preference vote and is not counted as a win.

Should you pick Gemma 4 31B or DeepSeek V4.1 Flash?

Pick Gemma 4 31B if…

  • You want the lighter model to host: about 21 GB at 4-bit against 381 GB.
  • Cost matters: $0.340 per million output tokens against $1.20.
  • You want to run it locally: at 4-bit it fits one 24–32 GB consumer GPU.
  • You want it running in minutes on a dedicated GPU through OpenLLM Buddy, in an uncensored build.

Pick DeepSeek V4.1 Flash if…

  • You care most about GPQA Diamond, Humanity's Last Exam, Humanity's Last Exam (with tools), where it scores at least 2 points higher.
  • People prefer its answers in blind chats: LMArena Elo 1476 against 1452.
  • You need fast responses: about 217 tokens per second through APIs, against 35.
  • You want it running in minutes on a dedicated GPU through OpenLLM Buddy, in an uncensored build.

Hardware and cost to run

Gemma 4 31B

At 4-bit it needs about 21 GB, which fits a single RTX 4090, so it can run on a home workstation with llama.cpp, Ollama, or vLLM. Full BF16 precision needs about 71 GB (1× H100 (80 GB)).

On OpenLLM Buddy: gemma-4-31B-it-uncensored runs on RTX 4090 24GB, with an OpenAI-compatible API and no per-token fees. Prices start at $13 USD for 11 hours.

Deploy Gemma 4 31B
Full GPU memory table for Gemma 4 31B

DeepSeek V4.1 Flash

Even at 4-bit it needs about 381 GB (4× H200 (564 GB)), well beyond a home workstation, so most people run it on rented multi-GPU servers. Full BF16 precision needs about 1.3 TB (Multiple nodes (over 1.1 TB)).

On OpenLLM Buddy: DeepSeek-V4.1-Flash-UNCENSORED-FP8 runs on 4× H200 141GB as an FP8 build, with an OpenAI-compatible API and no per-token fees. Prices start at $280 USD for 11 hours.

Deploy DeepSeek V4.1 Flash
Full GPU memory table for DeepSeek V4.1 Flash

Category profile

Category index uses every test each model reports, not only the shared ones.

ReasoningCodingAgenticKnowledge
Gemma 4 31BDeepSeek V4.1 Flash
Reasoning69.875.9
Coding85.672.1
Agentic68.991.3
Knowledge92.5—

Specs side by side

Gemma 4 31BDeepSeek V4.1 Flash
LabGoogleDeepSeek
WeightsOpenOpen
Released2026-04-022026-09-10
Parameters30.7B dense552B total, 16B active (MoE)
Context256K1M
LicenseApache-2.0MIT
4-bit memory21 GB · RTX 4090 (24 GB)381 GB · 4× H200 (564 GB)
API price$0.100 in / $0.340 out per 1M tokens$0.300 in / $1.20 out per 1M tokens
API speed35 tok/s217 tok/s
One-GPU speed (est.)≈ 25–34 tok/s on RTX 4090 (4-bit)—

Related comparisons

Sources

ModelTestScoreSettingSource
Gemma 4 31BGPQA Diamond84.3—Gemma 4 31B IT model card (Hugging Face)
DeepSeek V4.1 FlashGPQA Diamond90.9Pass@1, max reasoning effortDeepSeek-V4.1-Flash model card (Hugging Face)
Gemma 4 31BHumanity's Last Exam19.5no toolsGemma 4 31B IT model card (Hugging Face)
DeepSeek V4.1 FlashHumanity's Last Exam36.8Pass@1, no tools, max effort (text-only subset: 39.1)DeepSeek-V4.1-Flash model card (Hugging Face)
Gemma 4 31BHumanity's Last Exam (with tools)26.5with searchGemma 4 31B IT model card (Hugging Face)
DeepSeek V4.1 FlashHumanity's Last Exam (with tools)63.9Pass@1, with tools, max effortDeepSeek-V4.1-Flash model card (Hugging Face)
Gemma 4 31BLMArena Text Elo1452LMArena text overall, 'Gemma 4 31B', ±7, as of 2026-10-27Arena (LMArena) text leaderboard
DeepSeek V4.1 FlashLMArena Text Elo1476LMArena text overall, 'DeepSeek-V4.1-Flash-20260910 (Max)', ±6, as of 2026-10-27Arena (LMArena) text leaderboard

Scores are published results from model cards, lab reports, and public leaderboards, each linked to its source. Settings such as thinking mode or test version differ between sources and are shown next to each score. OpenLLM Buddy catalog models are uncensored fine-tunes; their scores are those of the official base model. Data collected 2026-10-30. How the index works