Qwen3.6-35B-A3B
- Reported parameters
- 35.95B
- Release license
- apache-2.0
Downloadable models · Side by side
Qwen3.6 35B-A3B vs Gemma 4 26B A4B
Qwen3.6 35B-A3B and Gemma 4 26B A4B are the mixture-of-experts models in their families: each keeps 26 to 36 billion parameters in memory but uses only 3 to 4 billion for each token, so per-token compute is closer to a small model than a large one. Both are Apache 2.0, both take text and images, and both support 256K-token context. They differ in memory, default thinking behavior and how far they are tuned for agentic coding.
gemma-4-26B-A4B-it
- Reported parameters
- 25.81B
- Release license
- apache-2.0
The short answer
Choose Gemma 4 26B A4B if memory is tight: about 14.9 GB at 4-bit for one 8K-token conversation against 20.3 GB for Qwen3.6, and it still fits a 24 GB card at its full 256K context. Choose Qwen3.6 35B-A3B for coding agents and tool-heavy work: Qwen built this release around agentic coding, and it can keep earlier reasoning across turns.
Compare the details
Compare context, supported inputs and usage costs. Highlighted rows show a difference.
| Feature | Qwen3.6-35B-A3B | gemma-4-26B-A4B-it |
|---|---|---|
| Input formats | text | text |
| Output formats | text | text |
| Parameter count | 35,951,822,704 | 25,805,936,206 |
| Release license | apache-2.0 | apache-2.0 |
| USD input / 1M tokens | $0.10 | $0.08 |
| USD output / 1M tokens | $0.90 | $0.26 |
| Model identifier | Qwen/Qwen3.6-35B-A3B | google/gemma-4-26B-A4B-it |
API pricing
Prices are in USD per million tokens. Input is what you send; output is what the model generates. Chat subscriptions and API usage are billed separately.
Tools, images and additional reasoning can add charges. The examples below cover input and output tokens only.
Compare local modelsBeyond the price tag
Memory
For one 8K-token conversation with a 20% allowance, Qwen3.6 35B-A3B needs about 20.3 GB at 4-bit, 40.4 GB at 8-bit and 80.6 GB at 16-bit. Gemma 4 26B A4B needs about 14.9 GB, 29.3 GB and 58.2 GB. The gap is almost all weights. Qwen uses fewer parameters per token (3B against 3.8B), but every expert has to sit in memory, and Qwen stores about 36 billion parameters against Gemma's 26 billion.
Long context
Both keep long-context memory low. Qwen3.6 uses full attention in only 10 of its 40 layers, with 2 key-value heads each; the other 30 use Gated DeltaNet linear attention with a fixed-size state. Gemma 4 uses full attention in 5 of its 30 layers, and the other 25 only look back 1,024 tokens. At 4-bit we estimate about 23.1 GB for Qwen at 128K tokens and 26.1 GB at 256K, against 17.7 GB and 20.7 GB for Gemma. Gemma runs its full context on a 24 GB GPU; Qwen fits 128K there with little room to spare and needs a larger card for 256K.
Thinking mode
Qwen3.6 thinks before every answer unless you set enable_thinking to false, and Qwen advises keeping at least 128K tokens of context to preserve that ability, which leaves little headroom on a 24 GB card. Its preserve_thinking option keeps reasoning from earlier turns, which Qwen says helps agents stay consistent and can cut repeated reasoning. Gemma 4 only thinks when the system prompt starts with its think token. With thinking off it answers directly, which means fewer output tokens per reply for chat and extraction work.
Coding and agents
In Qwen's own benchmark table, run with its internal agent scaffold, Qwen3.6 35B-A3B scores 73.4 on SWE-bench Verified and 51.5 on Terminal-Bench 2.0, against 17.4 and 34.2 for Gemma 4 26B A4B. These are vendor-run numbers, not independent tests, but the gap is large enough to make Qwen the first one to try for code agents. Both support function calling: Qwen through the qwen3_coder tool parser in vLLM and SGLang, Gemma through native tool-use tokens in its chat template.
Inputs and serving
Both read images and short video and return text; Gemma 4 handles video as frames, up to 60 seconds at one frame per second. Neither takes audio; in Gemma 4 that is limited to the E2B, E4B and 12B models. Qwen3.6 was trained with multi-token prediction, which vLLM and SGLang can use for speculative decoding to speed up generation. Qwen recommends vLLM 0.19.0 or SGLang 0.5.10 or newer, so check your inference server version before switching.
Common questions
How much VRAM do Qwen3.6 35B-A3B and Gemma 4 26B A4B need?
For one 8K-token conversation including a 20% allowance: Qwen3.6 35B-A3B needs about 20.3 GB at 4-bit, 40.4 GB at 8-bit and 80.6 GB at 16-bit; Gemma 4 26B A4B needs about 14.9 GB, 29.3 GB and 58.2 GB. A 24 GB GPU runs either at 4-bit, and a 48 GB GPU runs either at 8-bit.
Which one is faster?
Speed in mixture-of-experts models follows active parameters, not total size. Qwen3.6 uses about 3B parameters per token and Gemma 4 about 3.8B, so both are much faster than dense models of 26B to 35B on the same GPU. Qwen thinks by default, and the extra output tokens usually affect response time more than the small gap in active parameters.
Which has the longer context window?
Both support 262,144 tokens natively. Qwen3.6 can be extended to about 1,010,000 tokens with YaRN; Gemma 4 26B A4B stops at 256K. Gemma needs less memory for the same long context.
Can I use them commercially?
Yes. Both are released under Apache 2.0, which allows commercial use and modification.
Can I run them on a CPU-only server?
Yes, at 4-bit with about 22 GB of RAM for Qwen3.6 or 17 GB for Gemma 4, including the operating system. Because only 3 to 4 billion parameters work on each token, they are far more usable on a CPU than dense models of the same size, though still much slower than on a GPU.
Run it yourself: memory and hosting
Memory needed to load each model and hold one 8,192-token conversation, with a 20% allowance for runtime buffers. The GPU and server columns show the cheapest listed options in our catalog that fit.
| Model and precision | Weights | Context memory | Plan for | Cheapest GPU that fits |
|---|---|---|---|---|
| Qwen3.6-35B-A3B4-bit | 16.7 GB | 0.2 GB | 20.3 GB | NVIDIA RTX A5000 (24 GB) · $0.27/hr Runpod · about $197/mo always on |
| Qwen3.6-35B-A3B8-bit | 33.5 GB | 0.2 GB | 40.4 GB | NVIDIA A6000 (48 GB) · $0.59/hr Runpod · about $431/mo always on |
| Qwen3.6-35B-A3B16-bit | 67.0 GB | 0.2 GB | 80.6 GB | NVIDIA GH200 Grace Hopper Superchip (96 GB) · $1.99/hr Vultr · about $1,453/mo always on |
| gemma-4-26B-A4B-it4-bit | 12.0 GB | 0.4 GB | 14.9 GB | NVIDIA RTX A5000 (24 GB) · $0.27/hr Runpod · about $197/mo always on |
| gemma-4-26B-A4B-it8-bit | 24.0 GB | 0.4 GB | 29.3 GB | NVIDIA A6000 (48 GB) · $0.59/hr Runpod · about $431/mo always on |
| gemma-4-26B-A4B-it16-bit | 48.1 GB | 0.4 GB | 58.2 GB | NVIDIA V100S ×2 (32 GB) · $1.76/hr OVHcloud · about $1,285/mo always on |
Qwen3.6-35B-A3B on a CPU server
At 4-bit, plan for about 22.3 GB of RAM including the operating system. CPU inference is much slower than a GPU, so this suits testing, background jobs and low traffic.
- Cheapest available server that fits: OVHcloud VPS-4 (US), $23.37/mo (24 GB RAM)
- Size it for your context and users
gemma-4-26B-A4B-it on a CPU server
At 4-bit, plan for about 16.9 GB of RAM including the operating system. CPU inference is much slower than a GPU, so this suits testing, background jobs and low traffic.
- Cheapest available server that fits: OVHcloud VPS-4 (US), $23.37/mo (24 GB RAM)
- Size it for your context and users
Estimates, not measurements. 4-bit and 8-bit figures assume common quantized files (such as GGUF Q4/Q8); real files add a little overhead. Longer contexts and more simultaneous users need more memory.