Gemma 4 31B
A 31B-parameter open-weights language model from Google's Gemma 4 family.
HN commenters rate it among the best models to run at home12 and use it for local agentic coding and simple tasks like data extraction or trivial refactors.345 The main caveat is that it falls far short of frontier models like Sonnet or Opus on anything requiring real reasoning6, and one commenter finds Qwen 3.6 27B faster, smarter and more memory-efficient.7
Why people recommend it
Strong local model · 3
Named among the best models you can run at home, often alongside Qwen 3.6 27B.
“the best models you can run at home are probably Qwen 3.6 27B and Gemma 4 31B”
“Running llamacpp with Q4KM Qwen 3.6 27B or Gemma 4 31B”
Show 1 more
“It’s Qwen 3.6 27B with Gemma 4 31B, both unquantized, bf16/fp16”
Agentic coding · 2
One commenter runs it for all their agentic coding and finds it useful.
“I find it useful.”
“Gemma4-31b is working well for me on these types of tasks”
Simple tasks · 1
Used for easy work like data extraction and trivial refactors.
“or even to self-hosted Gemma4:31B”
Privacy · 1
Run locally when privacy matters.
“Running llamacpp with Q4KM Qwen 3.6 27B or Gemma 4 31B”
Creative writing · 1
Reported secondhand to be highly regarded for creative writing, especially its finetunes.
“Gemma4-31b seems to be very highly regarded for creative writing, especially its finetunes.”
Caveats and dissent
Not frontier-level · 1
A commenter who benchmarks local models says they are nowhere near Sonnet or Opus for tasks needing reasoning.
“they are NOWHERE near the huge models like Sonnet or Opus. Nowhere.”
Qwen often better · 1
One commenter says Qwen 3.6 27B is faster and smarter, and Gemma 4 31B is bigger, slower and less memory-efficient.
“I assume the 31B gemma has cases in which it edges out, I've just failed finding any”
Needs careful scoping · 1
Getting good coding results takes careful architecture and TDD.
“I find it useful.”
Often compared with
Who recommends it in their own words
Threads where it came up
- Apple Silicon costs more than OpenRouter 2 mentions
- Why isn't the industry freaking out about DeepSeek 4.1 Flash? 1 mention
- We Have AI at Home Chapter 1: A Box of Scraps 1 mention
- Show HN: OpenKnowledge – open-source alternative to Obsidian/Notion 1 mention
- Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding? 1 mention
- Ask HN: What does your local LLM setup looks like? 1 mention
Also written as Gemma4:31B, Gemma4-31b. Data: gemma-4-31b.json. Quotes are verbatim; the verdict and theme labels are written by a model from these comments, and every clause links to its sources.
All comments 9
“or even to self-hosted Gemma4:31B”
“the best models you can run at home are probably Qwen 3.6 27B and Gemma 4 31B”
“Gemma4-31b is working well for me on these types of tasks”
“I find it useful.”
“Running llamacpp with Q4KM Qwen 3.6 27B or Gemma 4 31B”
“Gemma4-31b seems to be very highly regarded for creative writing, especially its finetunes.”
“they are NOWHERE near the huge models like Sonnet or Opus. Nowhere.”
“I assume the 31B gemma has cases in which it edges out, I've just failed finding any”
“It’s Qwen 3.6 27B with Gemma 4 31B, both unquantized, bf16/fp16”