Gemma4-26B-A4B shows speed advantage over Qwen3.6-35B-A3B on Radeon 9070 XT

This v0 review assesses community feedback on Qwen3.6-35B-A3B and Gemma4-26B-A4B, focusing on reported performance with llama.cpp on AMD Radeon 9070 XT hardware.

TL;DR

Best for: Users prioritizing inference speed on AMD GPUs with llama.cpp, particularly the Radeon 9070 XT (Gemma4-26B-A4B). Skip if: You require detailed qualitative output comparisons, have different hardware, or need a larger model's potential capabilities. Bottom line: Gemma4-26B-A4B appears significantly faster on specific AMD hardware, while Qwen3.6-35B-A3B offers subjectively "nice results" at a slower pace.

METHODOLOGY

This v0 review draws on a single community report from Reddit user MarcCDB, published on May 24, 2026. Independent benchmarks are pending. Update cadence: re-tested when claims diverge from observed behavior or when more comprehensive community data emerges.

• Tool Name + Version + Date Observed: Qwen3.6-35B-A3B and Gemma4-26B-A4B, observed May 24, 2026.
• Source Signal URL: https://www.reddit.com/r/LocalLLaMA/comments/1tmbola/qwen3635ba3b_vs_gemma426ba4b/
• What's Covered in This Review: This review covers MarcCDB's anecdotal experience comparing the two models. Specifically, it addresses their reported performance on a Radeon 9070 XT GPU using the latest llama.cpp build. MarcCDB noted that Qwen provided "nice results" while Gemma4 ran "so much faster."
• What's NOT Covered: This review does not include independent performance benchmarks (e.g., tokens/second, memory usage), long-term workflow integration, or detailed qualitative comparisons of model output. Edge cases, specific task performance, and comparisons across different hardware or llama.cpp versions are also not covered.

WHAT IT DOES

Qwen3.6-35B-A3B: Alibaba's locally-run model

Qwen3.6-35B-A3B is a quantized version of Alibaba Cloud's Qwen large language model series, specifically the 3.6-35B parameter variant. The "A3B" suffix typically denotes a 3-bit quantization scheme, designed to reduce memory footprint and improve inference speed on consumer hardware. Users often deploy such models locally using inference engines like llama.cpp to leverage their own GPUs or CPUs. MarcCDB reported achieving "nice results" with Qwen, suggesting satisfactory output quality for their use cases.

Gemma4-26B-A4B: Google's efficient alternative

Gemma4-26B-A4B is a quantized version of Google's Gemma model, a lightweight, open model family derived from the Gemini research. The "26B" indicates a 26 billion parameter count, making it smaller than the Qwen variant under comparison. The "A4B" suffix points to a 4-bit quantization, which generally offers a balance between performance and model fidelity. MarcCDB's primary observation was that Gemma4 runs "so much faster" on their specific hardware configuration, highlighting its potential for high-speed local inference.

llama.cpp for local inference

Both models were run using llama.cpp, an inference engine optimized for running large language models on consumer hardware, including CPUs and various GPUs. Its continuous development aims to maximize performance across diverse systems, making it a common choice for local LLM experimentation and deployment.

WHAT'S INTERESTING / WHAT'S NOT

What's interesting here is the specific hardware context provided by MarcCDB: a Radeon 9070 XT GPU running the latest llama.cpp. This offers a concrete, albeit anecdotal, data point for users with similar AMD hardware. The reported speed advantage of Gemma4-26B-A4B over Qwen3.6-35B-A3B, despite the latter having more parameters (35B vs 26B) and a more aggressive quantization (3-bit vs 4-bit), suggests that architectural differences or llama.cpp's optimization for Gemma might play a significant role on this particular stack. This kind of real-world, hardware-specific feedback is invaluable for the local LLM community.

What's not interesting, or rather, what's missing, is the quantification of these observations. "Nice results" for Qwen and "so much faster" for Gemma lack the specific metrics needed for a robust comparison. Without tokens-per-second, memory usage figures, or examples of the "nice results," it is difficult to generalize these findings or understand the trade-offs involved. The absence of context regarding the types of tasks performed also limits the utility of the qualitative feedback. This signal highlights a common challenge in community-driven tool evaluation: enthusiasm often outpaces detailed, reproducible benchmarking.

PRICING

Both Qwen and Gemma models are open-source and available for free download and local inference. Users incur costs only for the hardware required to run them (e.g., GPU, CPU, RAM) and the electricity consumed. This review's pricing snapshot date is May 24, 2026.

VERDICT

Based on MarcCDB's experience, Gemma4-26B-A4B is the clear choice if inference speed on a Radeon 9070 XT with llama.cpp is your primary concern. Its reported "so much faster" performance outweighs Qwen3.6-35B-A3B's "nice results" when raw speed is paramount. While Qwen offers a larger parameter count, potentially leading to more nuanced or capable outputs, the performance delta observed on this specific hardware stack makes Gemma a more pragmatic option for local, high-throughput applications. For users with different hardware or those prioritizing output quality over speed, Qwen might still be a contender, but the current data points to Gemma for efficiency on AMD.

WHAT WE'D TEST NEXT

To provide a more comprehensive review, we would conduct controlled benchmarks focusing on several key areas. First, we would quantify the "so much faster" claim by measuring tokens-per-second for both models across various prompt lengths and complexities on the Radeon 9070 XT. Second, we would evaluate memory usage during inference to understand the true hardware requirements. Third, a qualitative assessment of output quality for common tasks (e.g., summarization, code generation, creative writing) would clarify what constitutes "nice results" for Qwen and how Gemma compares. Finally, we would expand testing to other popular AMD and NVIDIA GPUs, as well as different llama.cpp versions and quantization levels, to determine the generalizability of these findings.

Pull quote: “MarcCDB's primary observation was that Gemma4 runs "so much faster" on their specific hardware configuration, highlighting its potential for high-speed local inference.”