← All models

Gemma 4 26B A4B IT Assistant

Published Publisher Google

Positioning

An open-weight multi-token prediction (MTP) drafter for the instruction-tuned Gemma 4 26B A4B target model. It is an inference accelerator, not a standalone replacement for the target model. Its release identifier is google/gemma-4-26B-A4B-it-assistant. The model repository lists approximately 0.4B parameters for the assistant; the target model is a mixture-of-experts model with approximately 25.2B total and 3.8B active parameters. (ai.google.dev)

Capabilities and uses

The drafter proposes multiple candidate tokens, which the target model verifies. It uses the target model’s activations and shared KV cache. Google states that this speculative decoding preserves the target model’s output quality while potentially reducing generation latency. Applicable uses include responsive chat, coding assistants and agentic workflows when paired with the matching target model. (ai.google.dev)

Limitations

  • The assistant requires the matching target model and a compatible inference implementation; its capabilities should not be treated as an independent chat API.
  • Speedup depends on hardware, workload and batch size. Google notes that the 26B A4B drafter may not speed up batch-size-one inference on hardware without sufficient parallelism. (ai.google.dev)
  • The target model can still produce incorrect or outdated information and reflect biases; acceleration does not remove those risks. (huggingface.co)

Release and license

Google’s announcement is dated May 5, 2026 and says the MTP drafters were available under the Apache 2.0 license. Google’s separate release log lists Gemma 4 MTP for the 26B A4B family under April 16, 2026; the two official dates therefore differ. (blog.google)

Pricing

As of November 16, 2026, a separate official hosted input-token or output-token price for this assistant checkpoint could not be verified. Currency, rates and billing units are therefore unverified. Downloadable weights are available, but self-hosting may incur infrastructure costs; pricing for the separate target model should not be assumed to apply to this drafter. See Google’s Gemini API pricing page and model repository. (ai.google.dev)

Official sources