Can you run Gemma 4 26B-A4B on RTX 5090?

Gemma 4 26B-A4B Q4_K_M on the NVIDIA GeForce RTX 5090: verdict, VRAM math and estimated speed.

Yes, it runs
Quick answer

Yes, the NVIDIA GeForce RTX 5090 runs Gemma 4 26B-A4B. 16 GB weights at Q4_K_M vs 28.8 GB usable VRAM; ~118 tok/s est. (ModelFit, 2026).

$ollama run gemma4:26b
VERDICT
Comfortable
EST. SPEED
~118 tok/s
WEIGHTS
16 GB Q4_K_M

VRAM math and speed are ModelFit engine estimates, not measurements. Commands are registry-verified Ollama tags.

Cite this page: ModelFit, Gemma 4 26B-A4B on RTX 5090, https://modelfit.io/can-i-run/gemma4-26b-a4b-q4-on-rtx-5090/, updated September 2026, CC BY 4.0.

Last updated: September 18, 2026 · Editor: ModelFit Team

VRAM
32 GB (28.8 usable)
Model weights
16 GB Q4_K_M
Est. speed
~118 tok/s
First token
~0.3s
Fit grade
A · 44

What limits this combo

The four constraints the fit engine checks for Gemma 4 26B-A4B on the RTX 5090, in order of what usually breaks first.

Weights fitOK16 GB of weights against 28.8 GB usable — 12.8 GB of headroom for context and the runtime.
Context ceilingOKWeights + KV cache stay in budget up to 32k tokens (24.0 GB total). Quantizing the KV cache to q8_0 roughly halves the KV column and pushes this further.
SpeedOK~118 tok/s (est.) on 1792 GB/s of memory bandwidth — comfortable for interactive chat and agentic loops.
Use-case ceilingTightStrongest fit: Chat (A). Agentic coding caps at C — either the speed floor or the model's tuning is the limit; see the workload table for the per-case reason.

Quant explorer

Switch between the quality-gated builds tracked for this exact combo. Q4_K_M stays the default recommendation; heavier Q6/Q8 builds only appear when their exact Ollama tags are registry-verified.

ModelFit engine estimate
Verdict
Runs
Est. speed
~118 tok/s
First token
~0.3s · Instant
Safe context
32k
A · 44 Excellent headroom

Why no 2-bit builds: Q1/Q2-class quants can look attractive in a memory table, but their quality loss is large enough that ModelFit excludes them from rankings instead of inflating the catalog with junk options. Usable VRAM: 28.8 GB.

Workload verdicts

Gemma 4 26B-A4B on the RTX 5090, graded per use case from the model's registry-verified tuning and the engine's speed estimate for this exact combo.

ModelFit engine estimate
ChatA~118 tok/s est. vs ~20 needed for chat.
CodingA~118 tok/s est. vs ~15 needed for coding.
Agentic codingC~118 tok/s est. vs ~30 needed for agentic coding; not tuned for tool-calling loops.
ReasoningC~118 tok/s est. vs ~12 needed for reasoning; not a reasoning-tuned model.
RAGA~118 tok/s est. vs ~20 needed for rag.

Grades combine tag-verified tuning (a model not built for the workload caps at C) with the tok/s floor each workload needs to feel usable. A combo that partially offloads caps at C; one that does not fit is D everywhere.

Memory math: weights + context vs budget

Weights take 16 GB. Context costs extra KV-cache on top — this is where long-context sessions break on cards that technically fit the weights.

Weights 16 GBKV cache · 16k context 4.0 GBHeadroom 8.8 GBRuntime reserve 3.2 GB

The model leaves about 8.8 GB of the usable VRAM budget free after weights and 16k context.

ContextKV-cacheTotalFits
8k tokens2.0 GB18.0 GBFits
16k tokens4.0 GB20.0 GBFits
32k tokens8.0 GB24.0 GBFits
64k tokens16.0 GB32.0 GBOver
128k tokens32.0 GB48.0 GBOver

KV-cache figures assume an fp16 cache, the llama.cpp/Ollama default. Standard GQA models use a size-class estimate (8 KV heads x 128 head dim class); hybrid linear-attention models (Qwen3.5/3.6, Qwen3-Next) use the exact per-token cost from their published config, since only their sparse full-attention layers cache KV. A q8_0 KV cache roughly halves either figure. Estimates, not measurements.

See how fast it feels

A deterministic typing simulation for Gemma 4 26B-A4B: first token ~0.3s (instant prefill), then ~118 tokens/sec.

Start the simulation to preview the response pace.

ModelFit engine estimate, not a measured benchmark. Real speed varies with prompt length, thermals, runtime, and KV-cache settings.

Gemma 4 26B-A4B on RTX 5090: FAQ

Can the NVIDIA GeForce RTX 5090 run Gemma 4 26B-A4B?

Yes. Gemma 4 26B-A4B (Q4_K_M) loads in about 16 GB and the RTX 5090 offers 28.8 GB of usable VRAM, leaving headroom for context. Expect at roughly 118 tokens/sec (est.).

How much VRAM does Gemma 4 26B-A4B need?

About 16 GB for the weights at Q4_K_M, plus KV-cache for context: roughly 4.0 GB extra at 16k tokens. The RTX 5090 budget is 28.8 GB (32 GB x 90%).

What is the best quantization of Gemma 4 26B-A4B for the RTX 5090?

The Q8_0 build is the highest quality that fits (28.1 GB vs 28.8 GB usable). The Q4_K_M build at 16 GB leaves more room for long context.

What GPU do I need to run Gemma 4 26B-A4B comfortably?

The RTX 5090 already runs Gemma 4 26B-A4B comfortably. Larger cards only buy you longer context or a heavier quant.

How much context can Gemma 4 26B-A4B use on the RTX 5090?

Up to 32k tokens stay fully in VRAM (24.0 GB total with the KV cache). At 64k tokens the total reaches 32.0 GB and tips over the budget. Quantizing the KV cache to q8_0 roughly halves the KV column and buys back about one context tier.

Is Gemma 4 26B-A4B on the RTX 5090 fast enough for coding agents?

Barely on paper: the engine estimates ~118 tok/s against the ~30 tok/s an agentic loop needs to feel responsive, but Gemma 4 26B-A4B is not tuned for tool-calling, which caps its agentic grade — a coding-tuned model is the better pick here.