---
title: "Uno on Gemma 4 26B MoE: BroadNet study and DOI | BroadRouter"
description: "BroadNet trained the first Uno self-drafting adapter for a mixture-of-experts model, Gemma 4 26B A4B. What it measured, when self-drafting pays, and the DOI."
url: https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/
lang: en
updated: 2026-11-15
publisher: BroadRouter (a BroadNet site); API by BroadNet Inference, https://inference.broadnet.ai
---

# Uno on Gemma 4 26B: BroadNet's mixture-of-experts study

By the BroadNet Inference team · 3 min read · Published 2026-11-15 · Updated 2026-11-16

BroadNet's September 2026 preprint (DOI 10.5281/zenodo.22820511) trains the first Uno adapter for a mixture-of-experts model, Gemma 4 26B A4B, served at 4-bit. The adapter decodes 1.166 times faster than plain decoding on a 72-prompt production workload, behind two separate drafters at 1.474 and 1.553, but leads the production drafter at about 10,500 tokens of context.

## Key takeaways

-   Preprint on Zenodo, DOI 10.5281/zenodo.22820511, version 2.3, 18 September 2026, CC BY 4.0. Not peer-reviewed.
-   Target: Gemma 4 26B A4B, the base of BroadNet Gemma 4 VL 26B, served with 4-bit weights.
-   A rank-16 adapter trained on 965 rows raised accepted drafts from 42.4% to 48.9%.
-   Speed-up over plain decoding: 1.166x for the Uno adapter, 1.474x for a multi-token-prediction drafter, 1.553x for a block drafter.
-   At about 10,500 tokens of context the adapter led the production drafter, 1.089x against 1.011x.

**On this page**

1.  What was the question?
2.  How did BroadNet test it?
3.  What did it find?
4.  What does it mean in practice?
5.  What are the limits?

BroadNet’s second Uno preprint applies self-drafting to a [mixture-of-experts](https://broadrouter.com/glossary/mixture-of-experts/) model for the first time. Published on Zenodo on 18 September 2026 under DOI [10.5281/zenodo.22820511](https://doi.org/10.5281/zenodo.22820511), it serves Gemma 4 26B A4B at 4-bit and trains an Uno adapter for it. The target is the same base model as [BroadNet Gemma 4 VL 26B](https://broadrouter.com/models/gemma-4-vl-26b/) on BroadNet Inference.

It follows the [Uno in vLLM serving study](https://broadrouter.com/blog/uno-in-vllm-serving-study/), which tested a dense 8B model. The numbers below are BroadNet’s; we have not reproduced them.

## What was the question?

Speculative decoding drafts several tokens and verifies them in one pass. On a mixture-of-experts model there is an extra worry: draft rows can route to different experts, so drafting four tokens might read many more expert weights than generating one. The paper asks whether that “expert fan-out” makes self-drafting too expensive, and when self-drafting beats a separate drafter.

## How did BroadNet test it?

From the paper’s abstract and BroadNet’s methods line:

-   **Model:** Gemma 4 26B A4B, about 4B parameters active per token, with 4-bit weights.
-   **Hardware:** 24 GB-class GPUs, one sequence at a time.
-   **Workload:** 72 prompts from BroadNet’s production traffic; each speed-up is divided by a plain-decoding run in the same session.
-   **Long context:** a separate set of five prompts at about 10,500 tokens.
-   **Adapter:** rank 16, trained on 965 rows through the same 4-bit checkpoint it is served on, on a single 24 GB GPU, with no full-precision weights in the loop.

## What did it find?

| How Gemma 4 26B A4B was decoded | Speed-up over plain decoding |
| --- | --- |
| Plain decoding | 1.000 |
| Uno pilot adapter, width 4 | 1.166 |
| Multi-token-prediction drafter, width 2 | 1.474 |
| Block drafter, width 8 | 1.553 |
| Uno at about 10,500 tokens of context | 1.089 (MTP drafter: 1.011) |

Other results BroadNet reports:

-   **Training helped.** Accepted share rose from 42.4% to 48.9% over 1,500 updates, and a larger 2,000-row corpus, trained on the same recipe, accepted 0.49 more draft tokens per verify step than the matched anchor run.
-   **Expert fan-out is real but cheap.** Four draft rows touched 22 distinct experts against one row’s 8, yet the whole row-related cost was bounded by 0.53 steps in a cycle of about 2.4. BroadNet concludes that routing cost is “worth under a fifth of a cycle” on this class of hardware.
-   **Lossless.** BroadNet built a check for a profile whose plain output does not reproduce itself across sessions, and reports Uno “as close to plain as plain is to itself”.
-   **Five languages.** The paper gives what it calls the first measurement of this adapter’s acceptance across five languages.

## What does it mean in practice?

The paper’s own rule is simple: if a good small drafter exists for your model, use it; self-drafting earns its place at long context and on models that have no drafter. That matches what BroadNet says it runs: its catalogue lists Gemma 4 26B in production with a multi-token-prediction drafter (k=2), with “90%+ acceptance on structured prompts”.

How much does a drafter add in production? BroadNet’s catalogue reports that multi-token prediction added 48.8% on an RTX 3090, and its Gemma 4 VL 26B page gives about 136 tokens per second on a single stream with speculative decoding on. Both are BroadNet’s own measurements, on the serving configuration its model page describes, so your speed through the API also depends on load and on the length of your prompt and answer.

For buyers of an API, the useful part is the method, not the adapter. BroadNet says every model it serves goes through this arithmetic before it reaches a customer: measure the cost of one draft cycle, then pick the drafter that wins on the actual workload. The same discipline led it to [rule out EAGLE3](https://broadrouter.com/blog/models-broadnet-ruled-out/) for saturated serving.

## What are the limits?

-   **Single-stream only.** One sequence at a time, one run per arm. Batch serving behaves differently, as the dense study showed.
-   **One model.** Results are for Gemma 4 26B A4B at 4-bit on 24 GB-class cards.
-   **A preprint.** Version 2.3 on Zenodo, not peer-reviewed, with a companion evidence archive attached.

Release v0.3.0 of [vllm-uno](https://github.com/brntech/vllm-uno/releases/tag/v0.3.0) ships the serving profile, a container image and the validation record, so the numbers can be checked. To try the served model itself, BroadNet’s [browser playground](https://broadrouter.com/guides/try-broadnet-inference-playground/) runs Gemma 4 with no account.

Free tools for this

-   [OpenAI to BroadNet migration helper](https://broadrouter.com/tools/openai-to-broadnet-migration/) Paste OpenAI SDK code; get base URL, key and model edits, a diff and the checks that remain.
-   [max\_tokens calculator for reasoning models](https://broadrouter.com/tools/reasoning-token-budget/) Answer length + reasoning budget + margin = a max\_tokens that does not cut the answer off.
-   [JSON schema generator for LLM structured output](https://broadrouter.com/tools/json-schema-generator/) Sample JSON to strict schema, Pydantic, Zod and a ready request.
-   [LLM JSON output validator](https://broadrouter.com/tools/llm-json-output-validator/) Parse errors by line and column, safe repairs, cut-off detection and JSON Schema checks for model output.

[All free tools](https://broadrouter.com/tools/)

## Terms in this article

[Token](https://broadrouter.com/glossary/token/)

A token is the unit of text a language model reads and writes: a word, part of a word, a punctuation mark or a byte sequence produced by the model's tokenizer. LLM APIs count, limit and usually bill usage in tokens.

[Speculative decoding](https://broadrouter.com/glossary/speculative-decoding/)

Speculative decoding is an inference technique in which a small, fast draft model proposes several next tokens and the large target model verifies them in a single forward pass, keeping the accepted ones. It speeds up generation without changing the distribution of the output.

[Mixture of experts (MoE)](https://broadrouter.com/glossary/mixture-of-experts/)

Mixture of experts (MoE) is a neural network design in which a layer holds many expert sub-networks and a learned router sends each token to only a few of them. An MoE language model has a large total parameter count but uses a small active subset for each token.

[Tokens per second](https://broadrouter.com/glossary/tokens-per-second/)

Tokens per second (TPS) is the rate at which a language model generates output tokens, usually measured for one request after its first token arrives. It sets how fast a reply streams once it starts. Summed across all concurrent requests it becomes system throughput; the wait before the first token is measured separately.

[vLLM](https://broadrouter.com/glossary/vllm/)

vLLM is an open-source library for LLM inference and serving, first developed at UC Berkeley's Sky Computing Lab. It manages KV cache memory with PagedAttention, batches incoming requests continuously, supports many quantization formats and exposes an OpenAI-compatible HTTP server.

## Questions

### What is a mixture-of-experts model?

A model whose layers contain many expert sub-networks, of which only a few run for each token. Gemma 4 26B A4B has 26B parameters in total and about 4B active per token, which is why it is cheap to serve for its size.

### Why is the Uno adapter slower than the other drafters here?

The paper explains it as structural: a self-drafting cycle costs two forward passes of the full model, while a separate drafter replaces one of them with a small network. So a dedicated small drafter usually wins when one exists.

### When does self-drafting pay?

According to the paper, at long context, where the wider draft commits more tokens per cycle (1.089x against 1.011x at about 10,500 tokens), and for models that have no small drafter at all.

### Does BroadNet use this adapter for its Gemma 4 API?

BroadNet's model page says its Gemma 4 26B build is served with multi-token-prediction speculative decoding (k=2). The Uno adapter is published as a research artifact; BroadNet does not say it serves API traffic with it.

### Can I download the adapter?

Yes. BroadNet publishes it on Hugging Face as Broadnet/gemma-4-26B-A4B-uno-pilot-adapter, with the serving profile and container image in vllm-uno release v0.3.0.

## Keep reading

-   [Uno in vLLM: what BroadNet's serving study measured](https://broadrouter.com/blog/uno-in-vllm-serving-study/)
-   [BroadNet's GPU hardware and serving approach, as BroadNet describes it](https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/)
-   [Models BroadNet evaluated and ruled out, and why](https://broadrouter.com/blog/models-broadnet-ruled-out/)
-   [When a 1M-token context window is worth it (and when not)](https://broadrouter.com/blog/million-token-context-window-guide/)
-   [A private endpoint starts with a clear workload](https://broadrouter.com/blog/private-endpoint-planning/)
-   [Before fine-tuning: define the problem to solve](https://broadrouter.com/blog/fine-tuning-readiness/)

Try it on your real workload. New accounts start with $5 of credit for Gemma and transcription. DeepSeek, GLM and Qwen3.8 need a top-up.

[Get started](https://inference.broadnet.ai/account/?utm_source=broadrouter&utm_medium=referral&utm_campaign=blog-uno-mixture-of-experts-gemma-4-study&utm_content=page-cta) ·

**Sources**

-   [Al Zubaidi, Self-drafting with Uno on a mixture of experts: serving and training an adapter for Gemma 4 26B A4B, Zenodo preprint v2.3, 2026-09-18 (record fetched 2026-11-15)](https://doi.org/10.5281/zenodo.22820511)
-   [BroadNet blog, Uno for vLLM now runs mixture-of-experts models (17 September 2026; fetched 2026-11-15)](https://www.broadnet.ai/blog/uno-mixture-of-experts)
-   [BroadNet Research: key findings (fetched 2026-11-15)](https://www.broadnet.ai/research)
-   [vllm-uno release v0.3.0 (fetched 2026-11-15)](https://github.com/brntech/vllm-uno/releases/tag/v0.3.0)
-   [BroadNet Gemma 4 VL 26B model page: serving configuration (fetched 2026-11-15)](https://www.broadnet.ai/inference/models/gemma-4-vl-26b)
