Gemma 4: what developers actually say

We read every Hacker News comment from Nov 2025 – Nov 2026 that gave a real opinion on Gemma 4. 57 of the 68 opinions come from people who say they use it themselves.

62%approval, all opinions

–approval, last 90 days

#31of 45 models ranked

57different developers

Developers give Gemma 4 62% approval across 68 opinions, and most feedback comes from firsthand use. They praise its code quality (16 positive, 4 negative), speed (6 positive, 0 negative), and privacy and local use (6 positive, 0 negative). Reliability is the weak point (2 positive, 8 negative), with reports of loops and failed tasks after context compaction. Monthly approval moved from 64% in April to 45% in May, 70% in June and July, then 50% in August and September on only 4 and 3 opinions.

Pick it if

Want a local model for simple coding tasks and code overviews
Care about privacy and local deployment
Need fast responses or help finding security bugs

Look elsewhere if

Need dependable agentic coding on complex projects
Expect long sessions to survive context compaction
Need frontier-level design work on large codebases

Strengths and weaknesses

Each row counts opinions that mentioned the topic. Bars to the right of the line are praise, bars to the left are complaints.

PraiseComplaints

Code quality

Reliability

Privacy & local use

Autonomy

Speed

How opinion has moved

Monthly approval (line) and number of opinions (bars). Months with fewer than 3 opinions are left out.

What developers say

Praise

Reports a faster per-answer latency working with Gemma 4, but the patch does not work with Qwen 3.8 models.HN comment · Sep 19, 2026

Uses Gemma 4 a lot, building system prompts in OpenCode to improve its results.HN comment · Aug 12, 2026

Says Gemma 4 and Qwen 3.6 produce strong results when implementing plans handed over from Gemini.HN comment · Jul 12, 2026

Says local Gemma 4 under OpenCode handles tools well with good code quality, fine for simple projects.HN comment · Jul 9, 2026

Has used Gemma 4 for quick Python scripts and code overviews without worrying about needing a bigger model.HN comment · Jun 29, 2026

Says Gemma 4 31B clearly beats Qwen 3.6 27B at finding security bugs in their benchmarks.HN comment · Jun 29, 2026

Says Gemma 4 31B QAT beats Qwen 3.6 on their reasoning and C tests, especially when quantized.HN comment · Jun 29, 2026

Says Gemma 4 feels unusually smart for its size and rivals top models at finding security bugs.HN comment · Jun 28, 2026

Complaints

Runs Gemma 4 26B A4B locally and considers it fine for general needs but not for coding features.HN comment · Sep 1, 2026

Says Gemma 4 tasks fail after context compaction and loop, and it misfired commands during a Proxmox fix.HN comment · Aug 18, 2026

Says Gemma 4 12B feels like a downgrade from Qwen 2.5 14B for large projects.HN comment · Jun 30, 2026

Says local Gemma 4 falls well short of Opus and Fable on complex codebase design problems.HN comment · Jun 15, 2026

Found Gemma 4 8B hard to get to a useful state for coding tasks.HN comment · Jun 13, 2026

Says Gemma 4's MTP head occasionally broke markup in OpenCode, so they stopped using it.HN comment · Jun 12, 2026

Says Gemma 4 is notably worse at agentic coding than DeepSeek V4 Flash or Qwen 3.6 27B, despite token efficiency.HN comment · May 21, 2026

Says Gemma 4 keeps getting stuck in loops during agentic coding.HN comment · May 11, 2026

Summaries are paraphrased. Each links to the original comment.

Gemma 4 compared

Gemma 4 vs Qwen 3.6 Gemma 4 vs Qwen 3.5

FAQ

Is Gemma 4 worth it?

It depends on the job. It has 62% approval and positive marks for code quality, speed, and local use, but reliability is negative at 2 positive and 8 negative opinions.

Is Gemma 4 better than Qwen 3.6?

Opinions split. Six people preferred Gemma 4 over Qwen 3.6, while three preferred Qwen 3.6 and three preferred Qwen 3.5. One complaint calls Gemma 4 notably worse at agentic coding than DeepSeek V4 Flash or Qwen 3.6 27B.

Is Gemma 4 good for agentic coding?

Feedback is mixed. Autonomy has 3 positive and 3 negative opinions, and complaints mention loops, misfired commands, and failed tasks after context compaction. One developer says it is worse at agentic coding than DeepSeek V4 Flash or Qwen 3.6 27B.