What are the verified coding and reasoning benchmarks for Gemma 4 31B IT?

Gemma 4 31B IT achieves SWE-bench Verified: 48.5% (Standard), LiveCodeBench: 44.8%, AIME: 62%. Verified sources: Gemma 4 Technical Report (Google DeepMind).

Does Swarm isolate Gemma 4 31B IT code edits from other agents?

Yes. Swarm provisions isolated Git worktrees for every session and subagent, preventing file race conditions and keeping your main working tree clean until verification passes.

Can Gemma 4 31B IT be assigned to specific subagent roles?

Yes. In ~/.swarm/config.json, you can configure Gemma 4 31B IT as Primary/Planner (DAG task decomposition), Coder (implementation in isolated worktrees), or Finder (read-only symbol search).

Gemma 4 31B IT — Specs, Context & Capabilities

Q: What is the recommended Swarm multi-agent topology for Gemma 4 31B IT?

Swarm recommends Gemma 4 31B IT as Coder Subagent (Tier 2: Open-Weights Local Specialist). High-fidelity local coding engine for privacy-sensitive enterprise software teams. Best paired with: Gemini 3.8 Flash or Claude 3.7 Sonnet.

Q: How do I run Gemma 4 31B IT with Swarm?

Configure Gemma 4 31B IT via CLI using: swarm ctl model set --provider google --model google/gemma-4-31b-it, or switch models interactively inside the terminal TUI with /models.

Gemma 4 31B IT Benchmarks & Agentic Accuracy

Verified primary research & developer preference metrics

Primary Sources Verified

SWE-bench Verified

48.5%

StandardGoogle Open Models Report ↗

LiveCodeBench

44.8%

StandardLiveCodeBench ↗

LMSYS Coding Elo

1325Elo

Coding ArenaLMSYS Arena ↗

AIME 2024 / 2025

62%

StandardGoogle Open Models Report ↗

Pricing ($ / 1M tokens)

$0.2/ $0.4

$0.05 cached readStandard API Pricing

What This Means in Real Software Development

Open-weights model from Google DeepMind capable of running on local GPU clusters.
Scores 48.5% on SWE-bench Verified and 62.0% on AIME 2024.
Fits on high-end consumer workstations (24GB-32GB VRAM) with quantization.
100% private, zero-telemetry autonomous coding inside Swarm.

Swarm Multi-Agent Topology

Tier 2: Open-Weights Local SpecialistRole: Coder Subagent

Best Paired With: Gemini 3.8 Flash or Claude 3.7 Sonnet

High-fidelity local coding engine for privacy-sensitive enterprise software teams.

Git Worktree Isolation Advantage:

Executes local Git worktree refactors via vLLM or Ollama.

What Developers & Practitioners Are Saying

Real-world field observations, production strengths, and developer consensus

Field Consensus

"Google's open-weights model designed for local deployment and edge systems."

Verified Real-World Strengths

High code quality for a 31B parameter open-weights model
Can be run on local workstations with Ollama or vLLM
Clean instruction following without corporate preambles

Production Caveats & Gaps

Limited context window compared to cloud Gemini models

Primary Sources & Methodology Citations

Gemma 4 Technical Report — Google DeepMind (February 2026)

Gemma 4 31B IT model overview

Gemma 4 31B IT is listed in the public Swarm catalog through Google Gemini. This page summarizes the verified catalog fields used by Swarm without claiming real-time access for every provider account.

Provider: Google Gemini
Family: Gemma
Context window: 262.1K tokens
Thinking support: Not supported
Availability status: Available
Catalog ID: google/gemma-4-31b-it

Image input
Text input
Text output
Reasoning

Run Gemma 4 31B IT in Swarm

Verified CLI workflow & multi-agent architecture

Quickstart Guide →

bash — swarm run

$swarm ctl model set --provider google --model google/gemma-4-31b-it

# Or switch interactively inside the terminal TUI:

>/models

Subagent Role Matrix

Assign Gemma 4 31B IT to specialized system roles in ~/.swarm/config.json:

Primary / Planner: Orchestrates full task DAGs and multi-stage pipelines.
Coder Subagent: Generates code and unit tests inside private Git worktrees.
Finder Subagent: Performs read-only code search, symbol discovery, and mapping.

System Agent Roles →Why Swarm Architecture →