Skills2026-04-11 · source-backed

Benchmark Gemma 4 31B for document-heavy agentic tasks, Qwen 3.5 27B for code generation.

Issue

2026-04-11

Confidence

source-backed

Sources

Redaction

passed

Story

Head-to-head testing on consumer hardware (RTX 3090 Ti, 96GB RAM) shows Gemma 4 jumped from 13.5% to 66.4% on multi-needle retrieval, making its 256K context actually usable. Qwen 3.5 wins on MMLU-Pro, GPQA Diamond, and LiveCodeBench. Match the model to the task type.

Ollama supports Qwen / Shared entities / Same source domain / Shared topic / Earlier coverage
Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE
Linked by a graph relationship (Ollama supports Qwen); both cover Gemma, LocalLLaMA, Qwen, RTX; reported by the same outlet (reddit.com).
Ollama supports Qwen / Shared entities / Same source domain / Shared topic / What happened next / Tension
A 27B Model Just Matched the 400B One. It Fits on Your Laptop.
Linked by a graph relationship (Ollama supports Qwen); both cover LocalLLaMA, Qwen, RTX; reported by the same outlet (reddit.com).
Anthropic criticizes Qwen / Shared entities / Same source domain / Shared topic / What happened next / Tension
Kimi K2.6: A 1-Trillion-Parameter Open-Weights Model Just Beat GPT-5.4 on SWE-Bench Pro
Linked by a graph relationship (Anthropic criticizes Qwen); both cover Gemma, LocalLLaMA, Qwen; reported by the same outlet (reddit.com).
Qwen competes with Google / Shared entities / Same source domain / Shared topic / Earlier coverage
Gemma 4 31B Just Made Half Your API Budget Obsolete
Linked by a graph relationship (Qwen competes with Google); both cover Gemma, LiveCodeBench, LocalLLaMA; reported by the same outlet (reddit.com).
Ollama supports Qwen / Shared entities / Same source domain / Shared topic / Earlier coverage
Google Ships Gemma 4 Under Apache 2.0. Four Models, Arena #3, and the License That Changes Everything.
Linked by a graph relationship (Ollama supports Qwen); both cover Gemma, LocalLLaMA, Qwen; reported by the same outlet (reddit.com).
Qwen competes with Google / Shared entities / Shared topic / What happened next
Xiaomi Just Open-Sourced a 1-Trillion Parameter Model Under MIT License. It Built a Compiler in 4 Hours.
Linked by a graph relationship (Qwen competes with Google); both cover Gemma, LocalLLaMA, Qwen; overlapping topics (agentic, benchmark, context).
Alibaba released Qwen / Shared entities / Same source domain / Shared topic / What happened next
Unsloth Pegs Qwen3.8-27B's VRAM Need at 17GB
Linked by a graph relationship (Alibaba released Qwen); both cover LocalLLaMA, RAM; reported by the same outlet (reddit.com).
OpenCode uses Qwen / Shared entities / Shared topic / What happened next
Local models are good now, and the timing isn't a coincidence
Linked by a graph relationship (OpenCode uses Qwen); both cover Gemma, Qwen; overlapping topics (agentic, code, gemma).

Source trail

Head-to-head testing

Entities

Benchmark Gemma Gemma GPQA Diamond Head LiveCodeBench LocalLLaMA Match MMLU Qwen RAM RTX

Provenance

Canonical issue: Ramsay Research Agent — April 11, 2026
AI generated: no
Story unit: 2026-04-11-benchmark-gemma-4-31b-for-document-heavy-agentic-tasks-qwen-3-5-27b-for-code-generation
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

Benchmark Gemma 4 31B for document-heavy agentic tasks, Qwen 3.5 27B for code generation.

Story

Related stories

Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE

A 27B Model Just Matched the 400B One. It Fits on Your Laptop.

Kimi K2.6: A 1-Trillion-Parameter Open-Weights Model Just Beat GPT-5.4 on SWE-Bench Pro

Gemma 4 31B Just Made Half Your API Budget Obsolete

Google Ships Gemma 4 Under Apache 2.0. Four Models, Arena #3, and the License That Changes Everything.

Xiaomi Just Open-Sourced a 1-Trillion Parameter Model Under MIT License. It Built a Compiler in 4 Hours.

Unsloth Pegs Qwen3.8-27B's VRAM Need at 17GB

Local models are good now, and the timing isn't a coincidence

Source trail

Entities

Provenance

Benchmark Gemma 4 31B for document-heavy agentic tasks, Qwen 3.5 27B for code generation.

Story

Related stories

Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE

A 27B Model Just Matched the 400B One. It Fits on Your Laptop.

Kimi K2.6: A 1-Trillion-Parameter Open-Weights Model Just Beat GPT-5.4 on SWE-Bench Pro

Gemma 4 31B Just Made Half Your API Budget Obsolete

Google Ships Gemma 4 Under Apache 2.0. Four Models, Arena #3, and the License That Changes Everything.

Xiaomi Just Open-Sourced a 1-Trillion Parameter Model Under MIT License. It Built a Compiler in 4 Hours.

Unsloth Pegs Qwen3.8-27B's VRAM Need at 17GB

Local models are good now, and the timing isn't a coincidence

Source trail

Entities

Provenance