Command R 35B (Ollama) is Cohere's RAG and tool-use model, 19GB with a 128K token (131,072) context. Inside Atlas the pricing is Free (self-hosted), and roughly 21GB to serve means it fits a 24GB card. Command R 35B was designed around retrieval augmented generation and grounded citation, which maps directly onto how Atlas feeds retrieved code chunks into every turn. The license is the catch: Cohere's weights carry a non-production research license, so check the terms before deploying it commercially.
Why is Command R 35B (Ollama) a good fit for a retrieval-based coding agent?
Command R 35B (Ollama) is purpose-built for retrieval augmented generation and multi-step tool use, which is exactly the shape of an Atlas agent turn. Cohere designed the 19GB model around grounded citation, and Atlas feeds retrieved code chunks into every turn it runs.
Most models used in coding agents were trained to write code and then asked, after the fact, to also reason over retrieved context and call tools correctly. Command R 35B was built in the opposite order. Retrieval augmented generation and multi-step tool use are the design target, and an Atlas turn is precisely that: retrieve, reason over the retrieved chunk, call a permission-gated tool, read the result, go again. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and it indexes code by AST declarations using tree-sitter, not blind line windows, so what Command R 35B receives is a well-formed retrieval context rather than a pile of line ranges.
Does Command R 35B cite the code it used?
Command R 35B (Ollama) does grounded generation with citation behavior, so it tends to point at the retrieved chunk it used rather than inventing one. In an Atlas session over a 128K token (131,072) window, that behavior is the difference between a reviewable answer and a plausible guess.
Grounded citation is Command R 35B's signature behavior and it is underrated for agent work. When a model tells you which retrieved declaration it based a change on, you can check that declaration in seconds. When it does not, you are auditing a claim with no address. Atlas already supplies the mechanical audit trail, because Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and Atlas snapshots file changes as git patches so edits can be diffed and rolled back. Command R 35B adds the semantic half: not only what changed, but which piece of your code the model was actually looking at when it decided to change it.
What does the Cohere license mean for commercial use of Command R 35B?
Cohere's weights for Command R 35B carry a non-production research license. Review Cohere's license terms for the weights before using it on commercial work. Command R 35B is Free (self-hosted) to run, but free to run and free to deploy are different questions.
This is the most important paragraph on the page and it is not a technical one. Command R 35B is the most capable RAG-native model you can pull in 19GB, and the pricing is Free (self-hosted), which reads like an open invitation. The license does not match that reading. A non-production research license means an internal prototype, a benchmark, or a personal project is a different situation from shipping client work through it, and the terms are Cohere's to define, not this page's to summarize. Read them. If the answer is no, Mistral Small 24B at 14GB is Apache 2.0 licensed and can be deployed commercially with no additional negotiation, which is the natural fallback.
How much memory does Command R 35B need to run with Atlas?
Command R 35B (Ollama) is a 19GB pull and takes roughly 21GB to serve, so it fits a 24GB card. Registering it means adding command-r:35b to the atlas.json ollama provider with limit.context 131072 and limit.output 8192 for the full 128K window.
A 24GB card is the sweet spot Command R 35B was sized for, and getting a 128K token (131,072) context from 19GB of weights is a genuinely efficient trade at that tier. Pull it with ollama pull command-r:35b (19GB, aliased to :latest), set model to ollama/command-r:35b, and lean on Atlas hybrid search, semantic plus keyword, fused with reciprocal rank fusion, to feed it well-formed retrieval context. Generate that context with the local Ollama embedder, so a RAG-native model runs against a RAG index that never left your disk. That pairing is the point: a model built for retrieval, fed by retrieval you fully control.
When should you pick a different model than Command R 35B?
Pick a different model than Command R 35B (Ollama) for diff generation. Command R 35B is not a code specialist, so pair it with a coder for diff generation and let Command R handle retrieval-heavy reasoning. The 19GB is spent on RAG behavior, not on patch quality.
Command R 35B is a specialist, and the failure mode is treating it as a generalist. On a question like which module owns this behavior, why does this call path exist, or what will break if I change this interface, Command R 35B's retrieval training pays off. On writing the actual patch, a code specialist at a similar or smaller footprint is stronger. The good news is that Atlas does not force a choice: Atlas lets you switch the active model and provider on the fly with favorites and recents, and Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, so Command R 35B can own the retrieval-heavy planning phase and a coder tag can own the build.
Setup
- 01Pull it: ollama pull command-r:35b (19GB, aliased to :latest)
- 02Review Cohere's license terms for the weights before using it on commercial work
- 03Register command-r:35b in the atlas.json ollama provider with limit.context 131072 and limit.output 8192
- 04Set model to ollama/command-r:35b and lean on Atlas hybrid search (semantic plus keyword, fused with reciprocal rank fusion) to feed it well-formed retrieval context
- 05Generate that context with the local Ollama embedder, so a RAG-native model runs against a RAG index that never left your disk
Frequently asked questions
- can I use command r 35b commercially
- Check first. Cohere's weights carry a non-production research license, so review Cohere's license terms before using Command R 35B on commercial work, even though it is Free (self-hosted) to run.
- how much VRAM does command r 35b need
- Roughly 21GB to serve from 19GB of weights, so Command R 35B fits a 24GB card while still offering a 128K token (131,072) context.
- best local model for RAG and tool use
- Command R 35B is purpose-built for retrieval augmented generation and multi-step tool use, with grounded generation and citation behavior, which maps directly onto how Atlas feeds retrieved code chunks into every turn.
- does command r cite its sources
- Command R 35B does grounded generation with citation behavior, so it tends to point at the retrieved chunk it used rather than inventing one.
- command r 35b vs a code specialist model
- Command R 35B is not a code specialist. Pair it with a coder for diff generation and let Command R 35B handle retrieval-heavy reasoning over a 128K window.
- how do I add command-r to atlas.json
- Register command-r:35b in the atlas.json ollama provider with limit.context 131072 and limit.output 8192, then set model to ollama/command-r:35b after running ollama pull command-r:35b.
- how does atlas retrieve code for a RAG model
- Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and indexes code by AST declarations using tree-sitter, not blind line windows. Generate that context with the local Ollama embedder to keep it on your disk.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Julia: A Terminal-Native AI Coding Agent for Project.toml Packages in 2026
Atlas is a terminal-native AI coding agent for Julia in 2026. It reads dispatch signatures and Project.toml [deps], fixes type instabilities, runs Pkg.test(), and applies JuliaFormatter.
Atlas vs Crush: Terminal AI Coding Agents in 2026
Comparing Atlas and Crush, two terminal AI coding agents for developers in 2026. Atlas offers robust planning and diffing, while Crush features LSP integration and mid-session model switching.
Atlas vs JetBrains Junie in 2026: Terminal Agent vs IDE Debugger-Driving Agent
Atlas vs JetBrains Junie in 2026: Junie drives the IDE debugger and reports 61.6% resolved on SWE-Rebench, bundled from $10/mo. Atlas is a free, open source terminal agent with permission gates.
Atlas for SolidJS: A Terminal-Native AI Coding Agent Built on Solid, for Solid, in 2026
Atlas is a terminal-native AI coding agent for SolidJS in 2026. It reads your createSignal and createMemo graph, fixes destructured props that break reactivity, and runs vitest.
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 Next.js in 2026
Adopt Atlas, the terminal-native AI coding agent, for Next.js development in 2026. Enhance productivity across App Router, server components, and API routes with secure, reviewable AI assistance.
Atlas for Polars: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for Polars. Build LazyFrame chains, push scan_csv predicates into the reader, and read explain() plans in 2026.
Atlas for COBOL: Copybooks, PIC Clauses, and GnuCOBOL in 2026
Atlas is a terminal-native AI coding agent for COBOL in 2026. It reads your divisions and copybooks, explains a paragraph before touching it, and compiles with cobc under GnuCOBOL.