# Atlas with Gemma 4 26B A4B: The Open-Weights MoE Option in 2026 > Gemma 4 26B A4B is a sparse MoE with about 4B active parameters out of 26B total, carrying a 262,144 token context. Gemma 4 26B A4B is the sparse mixture of experts member of the Gemma 4 family: 26B total parameters with roughly 4B active per token, which is what the A4B suffix means and why it runs far faster than its total size suggests. Gemma 4 26B A4B carries a 262,144 token context and a 32,768 token output cap, is reasoning capable, and ships with open weights, so it can run entirely on your own hardware with no code leaving the machine. Pricing is not publicly listed in the registry, so hosted access depends on which provider you route through. ## Key takeaways - Gemma 4 26B A4B is a sparse MoE with about 4B active parameters out of 26B total, so throughput tracks a small dense model while capacity tracks a large one. - The context window is 262,144 tokens, double the 128K typical of open-weight models in this class, with a 32,768 token output cap. - Open weights and reasoning capability mean Gemma 4 26B A4B can run entirely on your own hardware with no code leaving the machine. - VRAM is set by the full 26B parameter count, not the roughly 4B active, so memory needs exceed a 4B dense model. - No public per-token price exists in the registry, so hosted cost depends on the provider you route through. ## What does the A4B in Gemma 4 26B A4B mean? The A4B suffix in Gemma 4 26B A4B means roughly 4B active parameters per token out of 26B total. Gemma 4 26B A4B is a sparse mixture of experts model, so its throughput tracks a small dense model while its capacity tracks a large one, which is why it runs far faster than 26B suggests. Mixture of experts routing is the whole story of Gemma 4 26B A4B. Each token activates only a slice of the network, about 4B parameters, so the compute per token is small. The other 22B of parameters are still there, holding capacity the router can reach for when a token needs it. For an Atlas session that means faster turns than a 26B dense model would give you, without the quality floor of an actual 4B dense model. The catch is memory, which is covered below. ## How big is the Gemma 4 26B A4B context window? Gemma 4 26B A4B carries a 262,144 token context, double the 128K window typical of open-weight models in its class, with a 32,768 token output cap. For Atlas that means a large slice of a repository fits in one session without a hosted frontier model. Context is usually where open-weight models disappoint. Gemma 4 26B A4B at 262,144 tokens is a real exception in the open-weight tier, and it pairs with Atlas's retrieval well. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and Atlas indexes code by AST declarations using tree-sitter rather than blind line windows, so the tokens that reach the model are dense declarations rather than padded line ranges. A 262,144 token window filled with AST-aligned code goes a long way, and the 32,768 token output cap is enough for a substantial unified diff. ## Can you run Gemma 4 26B A4B locally with Atlas and Ollama? Yes. Gemma 4 26B A4B is reasoning capable and open weight, so it can run entirely on your own hardware with no code leaving the machine. Atlas can also build its code index with local Ollama embeddings, which means both the model and the index stay off third-party servers. The local path in Atlas goes through an ollama block in atlas.json using "npm": "@ai-sdk/openai-compatible" with baseURL http://localhost:11434/v1. Set the models map limit to { "context": 262144, "output": 32768 } so Atlas knows the real ceilings for Gemma 4 26B A4B rather than guessing. Run `atlas device` to confirm Ollama and a GPU were detected, and verify the entry with `atlas models ollama` before pinning it as "model". Pull only tags you have confirmed exist on ollama.com. ## What VRAM does Gemma 4 26B A4B actually need? Memory footprint for Gemma 4 26B A4B is set by the full 26B parameter count even though only about 4B are active per token. VRAM needs are therefore higher than a 4B dense model, and that is the main practical cost of the mixture of experts architecture. Sparse activation buys compute, not memory. Every expert has to be resident for the router to choose it, so a machine that comfortably runs a 4B dense model will not necessarily run Gemma 4 26B A4B. Plan capacity against 26B total parameters, not against the roughly 4B active. Run `atlas device` to see whether Atlas detected a GPU and what it found before you commit to the local path, because the failure mode of an undersized machine is a session that swaps and crawls rather than one that refuses to start. ## What does Gemma 4 26B A4B cost to use through a provider? Gemma 4 26B A4B has no public per-token price in the registry, so hosted access depends on which provider you route through. For hosted use, export GEMINI_API_KEY and run `atlas models google` to see if gemma-4-26b-a4b-it resolves on your key. The absence of a listed price is a genuine planning problem, not a detail. With no per-token figure to model, you cannot forecast an Atlas bill the way you can with a Gemini Flash or Pro id. That pushes most serious Gemma 4 26B A4B usage toward self-hosting, where the cost is hardware rather than tokens and the open weights mean the model is genuinely yours. If you do route through a hosted provider, confirm the pricing on that provider's terms before you point Atlas's main "model" slot at it. ## When should you pick a different model than Gemma 4 26B A4B? Pick a different model than Gemma 4 26B A4B when you need a published per-token price or a context beyond 262,144 tokens. Gemma 4 26B A4B has no public price in the registry, and the Gemini Flash and Pro lines offer a 1,048,576 token window with clear pricing. Gemma 4 26B A4B is the right answer to one specific question: how do I run a capable, reasoning-enabled coding model without code leaving my machine. If that is not your constraint, a hosted model is usually simpler and cheaper to plan around. Atlas lets you switch the active model and provider on the fly with favorites and recents, so a local Gemma 4 26B A4B and a hosted checkpoint can coexist in the same switcher. Use the local one for sensitive repositories and the hosted one when raw capability matters more than data residency. ## Setup 1. For hosted access, export GEMINI_API_KEY and run `atlas models google` to see if gemma-4-26b-a4b-it resolves on your key. 2. For local use, run `ollama pull gemma3` style pulls only for tags you have confirmed exist on ollama.com, then check `atlas device` to see that Ollama and a GPU were detected. 3. Add an ollama block to atlas.json using "npm": "@ai-sdk/openai-compatible" with baseURL http://localhost:11434/v1. 4. Set the models map limit to { "context": 262144, "output": 32768 } to match the real ceilings. 5. Verify the entry with `atlas models ollama` before pinning it as "model". ## FAQ ### what does a4b mean in gemma 4 26b a4b The A4B suffix means roughly 4B active parameters per token out of 26B total. Gemma 4 26B A4B is a sparse mixture of experts model, so it runs far faster than its total parameter count suggests. ### how much vram does gemma 4 26b a4b need Memory footprint is set by the full 26B parameter count even though only about 4B parameters are active per token, so VRAM needs are higher than a 4B dense model. Run `atlas device` to check that Atlas detected Ollama and a GPU. ### what is the context window of gemma 4 26b a4b Gemma 4 26B A4B has a 262,144 token context window, double the 128K window typical of open-weight models in its class, plus a 32,768 token output cap. ### how much does gemma 4 26b a4b cost per token Gemma 4 26B A4B has no public per-token price in the registry, so hosted access cost depends on which provider you route through. Self-hosting the open weights turns the cost into hardware rather than tokens. ### how do i run gemma 4 26b a4b locally with atlas Add an ollama block to atlas.json using "npm": "@ai-sdk/openai-compatible" with baseURL http://localhost:11434/v1, set the models map limit to { "context": 262144, "output": 32768 }, and verify with `atlas models ollama` before pinning it as "model". ### can i use an open weights model to keep code off third party servers Yes. Gemma 4 26B A4B is open weight and reasoning capable, so it can run entirely on your own hardware, and Atlas can build its code index with local Ollama embeddings so the index stays local too. ### is gemma 4 26b a4b good for coding in atlas Gemma 4 26B A4B is reasoning capable with a 262,144 token context and a 32,768 token output cap, which is enough to read a large slice of a repository and emit a substantial unified diff for Atlas to render. --- Canonical HTML: https://runatlas.sh/resources/models/gemma-4-26b-a4b Source of truth: aeo_pages row `/resources/models/gemma-4-26b-a4b` (segment: Models) (this file is generated from it, never hand-edited). Licence: Atlas is proprietary with a free core. It is not open source and there is no public source repository.