OpenCoder 8B (Ollama) is the model you drive Atlas with when reproducibility matters more than raw capability. It is a fully open code LLM with open training data and an open training recipe, not just open weights, and it is Free (self-hosted). The download is 4.7GB and serves in roughly 7GB, so it fits an 8GB consumer GPU. The catch is the context window: 8K tokens (8,192), a hard ceiling for agentic loops that inject tool output every turn.
What is OpenCoder 8B good at inside Atlas?
OpenCoder 8B (Ollama) is best inside Atlas at focused, single-file code work on a machine where every layer must be auditable. The 4.7GB download serves in roughly 7GB, and the reason to pick it over a stronger 8B is reproducibility: you can audit how it was built.
OpenCoder 8B publishes its open training data and its training recipe, not just open weights, which is a materially different transparency claim from most models on Ollama. Inside Atlas that pairs naturally with the parts of the agent that are already inspectable. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so an OpenCoder-authored patch is never applied silently. Every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, which is the right posture for an 8B model whose reasoning you want to check. OpenCoder also does native English and Chinese chat, which is useful on bilingual codebases where comments and identifiers are mixed and an English-only model loses the thread halfway through a file.
How do the 8K context and the free price change how you use Atlas?
OpenCoder 8B (Ollama) costs nothing, Free (self-hosted), but it charges you in window: 8K tokens (8,192) is the entire budget for system prompt, retrieved code, tool output, and the reply. Ollama's 1.5b tag shows only a 4K window, so there is no smaller variant that preserves even this.
Running Atlas at 8K means the prompt must be curated rather than dumped. Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so what lands in an OpenCoder prompt is a whole function or a whole class rather than an arbitrary 60-line slice. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, which raises the odds that the handful of chunks you can afford at 8K are the right ones. The practical configuration is to feed retrieved AST chunks rather than whole files and let compaction handle long sessions. Because OpenCoder 8B is Free (self-hosted), the failure mode is a truncated session, never a surprise bill, which is a very different risk profile from a metered API.
Can OpenCoder 8B keep an entire Atlas workflow offline?
Yes. OpenCoder 8B (Ollama) plus the Atlas Ollama embedder keeps generation and retrieval on the same machine. Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers, so an audit-friendly model is paired with an audit-friendly, fully local retrieval path.
The argument for OpenCoder 8B is that you can inspect the training data. That argument gets weaker if your repository is being shipped to a third-party embedding API every time the index rebuilds. Using the Ollama embedder closes that gap: the vectors are produced by the same local runtime that serves opencoder:8b, and nothing about the repo is transmitted during indexing. Atlas snapshots file changes as git patches so edits can be diffed and rolled back, which means an OpenCoder patch that goes wrong is a local revert rather than an incident. For teams whose interest in OpenCoder is a compliance interest rather than a benchmark interest, the local embedder is not an optional extra, it is the other half of the claim.
When should you pick a different model instead of OpenCoder 8B?
Pick a different model when your Atlas sessions are long agentic loops. OpenCoder 8B (Ollama) caps at 8K tokens (8,192), and a build agent that reads files, runs commands, and re-reads output will exhaust that in a handful of turns. Reproducibility is the reason to accept the ceiling, not a workaround for it.
OpenCoder 8B is the right pick when the open recipe is the requirement: research reproduction, a compliance review that asks what the model was trained on, or a bilingual codebase where English and Chinese comments coexist. OpenCoder 8B is the wrong pick when you want a long-running Atlas build agent grinding through a multi-file refactor, because 8K tokens (8,192) forces compaction constantly. Atlas lets you switch the active model and provider on the fly with favorites and recents, so this is not a permanent commitment. A common shape is to keep opencoder:8b as one favorite for auditable work and a longer-window tag as another, switching per task rather than per project.
Does OpenCoder 8B run on an 8GB GPU?
OpenCoder 8B (Ollama) is a 4.7GB download that serves in roughly 7GB, so it runs on an 8GB consumer GPU. That is the sizing headline: a fully open code LLM, with open data and an open training recipe, on hardware that a single developer already owns, at a price of Free (self-hosted).
The 7GB figure covers the served weights at a modest context. Because OpenCoder 8B caps at 8K tokens (8,192), the KV cache never grows into the multi-gigabyte territory that long-window models reach, which is one of the few upside consequences of the small window: memory planning stays simple. On an 8GB card you can serve opencoder:8b and still run the Atlas TUI, which is a terminal-native TUI rendered with SolidJS through the OpenTUI renderer and costs almost nothing in comparison. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, so be aware that parallel subagents on a single 8GB card will contend for the same served model.
Setup
- 01Pull it: ollama pull opencoder:8b (4.7GB, aliased to :latest).
- 02Add opencoder:8b to the ollama provider models map in atlas.json with limit.context 8192 and limit.output 4096.
- 03With only 8K of window, configure Atlas to feed retrieved AST chunks rather than whole files, and let compaction handle long sessions.
- 04Confirm with atlas models ollama and set model to ollama/opencoder:8b.
- 05Use the Ollama embedder for the index so an audit-friendly model is paired with an audit-friendly, fully local retrieval path.
Frequently asked questions
- how to use opencoder 8b with atlas
- Run ollama pull opencoder:8b (4.7GB), add opencoder:8b to the ollama provider models map in atlas.json with limit.context 8192 and limit.output 4096, then confirm with atlas models ollama and set model to ollama/opencoder:8b.
- what is opencoder 8b's context window
- OpenCoder 8B (Ollama) has a context window of 8K tokens (8,192). That is a hard ceiling for agentic loops that inject tool output every turn, so Atlas should feed it retrieved AST chunks rather than whole files.
- how much does opencoder 8b cost to run
- OpenCoder 8B is Free (self-hosted). The only cost is hardware: a 4.7GB download that serves in roughly 7GB, which runs on an 8GB consumer GPU.
- is opencoder actually open source or just open weights
- OpenCoder 8B publishes open training data and an open training recipe, not just open weights, which is a materially different transparency claim and the main reason to pick it over a stronger 8B model.
- can opencoder 8b run on an 8gb gpu
- Yes. OpenCoder 8B is a 4.7GB download that serves in roughly 7GB, so it fits an 8GB consumer GPU. The 8K tokens (8,192) window keeps the KV cache small, which helps memory planning.
- does opencoder 8b work on chinese codebases
- OpenCoder 8B supports native English and Chinese chat, which is useful on bilingual codebases where comments and identifiers are mixed. Atlas indexes code by AST declarations using tree-sitter, so those chunks arrive intact.
- is there a smaller opencoder model with the same context
- No. Ollama's opencoder 1.5b tag shows only a 4K window, so there is no small variant that keeps even the 8K tokens (8,192) of opencoder:8b.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Crystal: A Terminal-Native AI Coding Agent for shard.yml Projects in 2026
Atlas is a terminal-native AI coding agent for Crystal in 2026. It reads shard.yml and union types, narrows Nil unions, runs crystal spec behind a prompt, and formats the diff.
Debug a Single Failing Test with Atlas in 2026
How to debug one failing test with Atlas in 2026: run it in isolation with bash, walk the call graph with the lsp tool, and fix the code, not the assertion.
Atlas for OCaml: A Terminal-Native AI Coding Agent for dune and opam Projects in 2026
Atlas is a terminal-native AI coding agent for OCaml in 2026. It reads dune stanzas and .mli signatures, runs dune runtest behind a prompt, and finishes with ocamlformat.
Run Atlas Headless in CI with Atlas (2026 Workflow)
How to run Atlas headless in CI in 2026: atlas run sends one prompt and exits when the session goes idle, with --format json, --command, and --continue for pipeline steps.
Run the Test Suite and Triage the Failures with Atlas in 2026
How to triage a failing test suite with Atlas in 2026: bash truncates at 2000 lines or 50 KB and saves the full log, then grep groups failures by root cause.
Atlas for Perl: A Terminal-Native AI Coding Agent for CPAN Distributions in 2026
Atlas is a terminal-native AI coding agent for Perl in 2026. It reads cpanfile deps and @EXPORT lists, writes Test2::V0 cases, runs prove -lr t/, and runs perltidy on the diff.
Atlas for .NET: Your AI Coding Agent in 2026
Atlas empowers .NET developers in 2026 with a terminal-native AI coding agent. Securely integrate Atlas with ASP.NET Core for web APIs and services, leveraging local embeddings and robust safety features for efficient
Audit a Repo with Parallel Subagents in Atlas (2026 Workflow)
How to audit a repo with parallel subagents in Atlas in 2026: the task tool launches explore subagents in their own sessions, so only conclusions return to your context.