Gemma 3 12B Instruct is the mid-size Gemma 3 instruct model on Amazon Bedrock, at $0.05 per Mtok input and $0.10 per Mtok output. It gives up the 27B model's extra capacity in exchange for less than half the price and a 131,072 token window. Inside Atlas, Gemma 3 12B Instruct is best used in the small_model slot, where its exactly 2x input-to-output pricing is easy to budget and its 8,192 token output ceiling is not a constraint. It has no reasoning mode, so it needs tighter prompts and explicit plan steps.
How much cheaper is Gemma 3 12B Instruct than the 27B model?
Gemma 3 12B Instruct costs $0.05 per Mtok input, less than half the 27B variant's $0.12, with exactly 2x input-to-output pricing at $0.10 per Mtok output. On Amazon Bedrock that clean 2x ratio makes Atlas spend trivial to forecast.
Budgeting an agent is easier when the ratio is simple. With Gemma 3 12B Instruct, every million output tokens costs exactly twice what a million input tokens cost, so a rough read-to-write ratio for your Atlas sessions converts directly into a dollar figure. What you give up is capacity: 12B dense parameters carry less than the 27B variant, and the difference shows on harder code. The saving is real, and so is the capability drop, which is why the small_model slot is the right home for gemma-3-12b-it.
What is the context window of Gemma 3 12B Instruct?
Gemma 3 12B Instruct has a 131,072 token context, which is 35 percent smaller than the 27B variant's 202,752 tokens. For Atlas that means less of a repository fits at once, so retrieval quality matters more than it does on a larger window.
A 131,072 token window is workable because Atlas is deliberate about what it puts in it. 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 context handed to Gemma 3 12B Instruct is already the relevant declarations. Still, the 35 percent gap against the 27B model's 202,752 tokens is a real limit on how wide a change Gemma 3 12B Instruct can reason about in one pass.
What can Gemma 3 12B Instruct output in a single Atlas turn?
Gemma 3 12B Instruct has an 8,192 token output ceiling, double the 4,096 tokens the 4B Gemma 3 allows on Bedrock. In Atlas that headroom is enough for a real unified diff on a small number of files, which is more than the 4B model can manage.
The output ceiling is what separates a model that can propose a change from one that can only describe it. At 8,192 tokens, Gemma 3 12B Instruct can emit a genuine diff, and Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so you see exactly what it produced. It is not enough for a sweeping refactor, but it is enough for the scoped, mechanical edits that a small_model should be handling in the first place.
How do you configure Gemma 3 12B Instruct in Atlas on Bedrock?
Export AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, and AWS_REGION (AWS_BEARER_TOKEN_BEDROCK also works), enable Gemma 3 12B access in the Bedrock model access page for your region, then run `atlas models amazon-bedrock` and confirm google.gemma-3-12b-it appears.
With the id confirmed, set "small_model": "amazon-bedrock/google.gemma-3-12b-it" in atlas.json so cheap background calls route to Gemma 3 12B Instruct, and leave "model" on a stronger id. Then compare with /models in the TUI. Bedrock model access is granted per region, so an id that resolves in one AWS_REGION may be absent in another, which is the most common reason google.gemma-3-12b-it fails to show up in the list.
Can you self-host Gemma 3 12B Instruct instead of paying Bedrock?
Gemma 3 12B Instruct has 12B dense weights, small enough to self-host on a single consumer GPU if you later want to leave Bedrock. The open-weight path means the model you rent at $0.05 per Mtok input is the same model you can run yourself.
Self-hosting changes the economics from per-token to per-machine, and it changes the privacy posture entirely. Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers, so a locally hosted Gemma 3 12B Instruct plus a local index gives you an Atlas setup where nothing leaves the machine. The 12B dense size is the enabling fact here: a 27B model is harder to fit on a single consumer GPU, and a 12B one usually is not.
When should you pick a different model than Gemma 3 12B Instruct?
Pick a different model than Gemma 3 12B Instruct when the work needs reasoning or a window larger than 131,072 tokens. Gemma 3 12B Instruct has no reasoning mode, so it needs tighter prompts and more explicit plan steps from Atlas than a reasoning-capable checkpoint does.
Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and that plan agent is where the missing reasoning has to be compensated for. If the plan is explicit and each step is small, Gemma 3 12B Instruct executes fine. If the task requires figuring out what the steps are, move up. Atlas lets you switch the active model and provider on the fly with favorites and recents, so keep gemma-3-12b-it in the small_model slot and put the thinking on a stronger id.
Setup
- 01Export AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, and AWS_REGION (AWS_BEARER_TOKEN_BEDROCK also works).
- 02Enable Gemma 3 12B access in the Bedrock model access page for your region.
- 03Run `atlas models amazon-bedrock` and confirm google.gemma-3-12b-it appears.
- 04Set "small_model": "amazon-bedrock/google.gemma-3-12b-it" so cheap background calls route here.
- 05Leave "model" on a stronger id and compare with /models in the TUI.
Frequently asked questions
- how much does gemma 3 12b cost on bedrock
- Gemma 3 12B Instruct costs $0.05 per Mtok input and $0.10 per Mtok output via Amazon Bedrock, an exactly 2x input-to-output ratio. That is less than half the 27B variant's $0.12 per Mtok input.
- what is the context window of gemma 3 12b instruct
- Gemma 3 12B Instruct has a 131,072 token context window, which is 35 percent smaller than the 27B variant's 202,752 tokens, plus an 8,192 token output ceiling.
- gemma 3 12b vs gemma 3 27b which should i use
- Gemma 3 12B Instruct costs $0.05 per Mtok input against the 27B model's $0.12 and has a 131,072 token window against 202,752. Use the 12B in Atlas's small_model slot and the 27B when capacity matters more than price.
- can i run gemma 3 12b on a single gpu
- Yes. 12B dense weights are small enough to self-host on a single consumer GPU if you later want to leave Bedrock, and the open weights are the same model you rent at $0.05 per Mtok input.
- how do i set gemma 3 12b as the small model in atlas
- After confirming google.gemma-3-12b-it with `atlas models amazon-bedrock`, set "small_model": "amazon-bedrock/google.gemma-3-12b-it" in atlas.json so cheap background calls route there, and leave "model" on a stronger id.
- google.gemma-3-12b-it not showing in atlas models list
- Bedrock model access is granted per region. Enable Gemma 3 12B access in the Bedrock model access page for your AWS_REGION, then re-run `atlas models amazon-bedrock` to confirm google.gemma-3-12b-it appears.
- does gemma 3 12b support reasoning
- No. Gemma 3 12B Instruct has no reasoning mode, so it needs tighter prompts and more explicit plan steps. Use the Atlas read-only plan agent to do the decomposition before the build agent runs.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Automate GitHub Issue and Pull Request Triage with Atlas (2026 Workflow)
How to automate GitHub issue and pull request triage with Atlas in 2026: the atlas github command checks the actor has admin or write permission before it does anything.
Atlas for dbt: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for dbt. Read the ref() DAG, convert a table model to incremental, run dbt build against dev, and add tests in 2026.
Atlas for C in 2026
Atlas is a terminal-native AI coding agent for C in 2026. Run it in a project with a Makefile, have it find memory leaks or add Unity tests, and review the diff.
Atlas vs Cursor: terminal AI coding agents compared (2026)
A grounded 2026 comparison of Atlas and Cursor across workflow, change review, extensibility, and pricing for developers choosing an AI coding agent.
Atlas vs OpenAI Codex CLI: Terminal AI Coding Agents in 2026
Comparing Atlas and OpenAI Codex CLI for terminal AI coding in 2026. Atlas offers a TUI, permission-gated tools, and BYO models, while OpenAI Codex CLI provides sandboxed execution and OpenAI model integration.
Atlas for Haskell in 2026
Atlas is a terminal-native AI coding agent for Haskell in 2026. Run it in a project with a .cabal file or a stack.yaml, let cabal build drive it, and review the diff.
Atlas vs Base44: Terminal AI Coding Agents in 2026
Compare Atlas, the terminal-native AI coding agent, with Base44, the Wix-owned no-code app builder, for developers in 2026. Evaluate features, pricing, and workflow.
Atlas for Nuxt: Auto-Imports, useAsyncData, and Nitro Handlers in 2026
Atlas is a terminal-native AI coding agent for Nuxt in 2026. It reads nuxt.config.ts, pages/ routes, composables/ auto-imports, and server/api/ Nitro handlers, and tests with @nuxt/test-utils.