← All models

Gemma 4 31B IT Assistant

Published Publisher Google

Overview

Gemma 4 31B IT Assistant is a lightweight multi-token prediction (MTP) drafter for speculative decoding with the instruction-tuned Gemma 4 31B target model. Its release repository is google/gemma-4-31B-it-assistant; this is a repository identifier, not a verified API model ID. The assistant has approximately 0.5 billion parameters, while the paired target model has approximately 30.7 billion parameters. (blog.google)

Capabilities and uses

  • The drafter proposes candidate tokens, which the target model verifies. Google reports up to 3× faster generation for Gemma 4 MTP drafters in tested configurations, without degradation in output quality or reasoning; actual gains depend on the workload, hardware, and runtime. (blog.google)
  • When paired with google/gemma-4-31B-it, it can accelerate applications such as chat, coding assistance, and agentic workflows. The target supports text and image input, video as frames, text output, function calling, and a context window of up to 256K tokens. These are target-model capabilities, not standalone capabilities established for the drafter. (blog.google)
  • Google provides downloadable weights under Apache 2.0 and describes use with compatible runtimes including Transformers, MLX, vLLM, SGLang, and Ollama. (blog.google)

Limitations

The drafter is intended to be paired with its corresponding target model, rather than treated as a standalone 31B assistant. Speedup is not guaranteed to reach the reported maximum. The 31B target does not support native audio input. Generated answers can be inaccurate, outdated, or biased and may require verification and safeguards. (blog.google)

Release date

May 5, 2026, the date of Google's Gemma 4 MTP drafter announcement. (blog.google)

Pricing

As of November 16, 2026, a separate official hosted API price for this assistant's input or output tokens, including its billing currency and unit, could not be verified. The weights are downloadable, but self-hosting may incur infrastructure costs. Pricing for a target model or another hosted service should not be assumed to apply to this drafter. Google's Agent Platform pricing page was checked; it does not establish a separate price for this assistant. (blog.google)

Official sources