Skip to main content
Gemma 4
Large tier
Tool calling

Gemma 4 26B-A4B

Best speed/quality MoE — 26B total, ~4B active, 256K context.

by Google

Gemma 4 group

Overview

Gemma 4 26B-A4B is the MoE instruct checkpoint for computer-use style jobs: strong quality with active-param serve cost closer to a mid model. Prefer over 31B when latency and VRAM matter more than peak scores.

Gemma 4 26B-A4B trains on your agent's own traces with LoRA adapters or full finetuning, is benchmarked against your production model before any traffic moves, and serves through the same OpenAI-compatible API as every Overmind model. The final weights are yours to download and run anywhere.

Good for

  • Agentic computer-use and long-context coding
  • QLoRA when dense 31B is too heavy
  • Multimodal text+image production agents

More in Gemma 4

← Back to all models
Train now

Specs & cost

Cost to trainfrom $2.00
Cost to run (1M tok)from $8.00
GroupGemma 4
FamilyGoogle
Parameters26B (4B active)
TierLarge
Context window256K
Max training context256K
Tool callingYes
Methods—
Train now
Training docs

On Overmind

  1. 01 Train on your data

    LoRA or full finetuning on a training-intent dataset from your agent's traces. Overmind recommends tiers from your data stats; you can edit every setting.

  2. 02 Benchmark vs production

    Each finished model is scored on your eval set with your rubric against the model your agent runs today. The headline is the baseline delta per metric.

  3. 03 Serve and own the weights

    Succeeded runs deploy automatically to Inference through the same OpenAI-compatible API. The final weights are yours to download and run anywhere.