DeepSeek-R1 (local via Ollama) is the open reasoning model that made visible chain-of-thought mainstream. Ollama ships it from 1.5B to 671B, so the same model name scales from a laptop to a server rack, with a 128K to 164K token context depending on the variant. It is free to self-host, or $1.35 / $5.40 per Mtok via Bedrock. Inside Atlas, DeepSeek-R1 is the model you reach for when a bug will not yield to a non-reasoning model.
What sizes does DeepSeek-R1 come in on Ollama?
DeepSeek-R1 is available at 1.5B, 7B, 8B, 14B, 32B, 70B, and 671B under a single `ollama pull deepseek-r1` tag family, so you can right-size to your hardware without changing prompts. That range, laptop to server rack, is unmatched by any other local reasoning model.
The single-tag family is the operational advantage of DeepSeek-R1. A team can standardize on one model name and let each machine pull the size it can serve: `ollama pull deepseek-r1:14b` for a 16GB card, or `ollama pull deepseek-r1:32b` for 24GB. The prompts, the Atlas configuration, and the agent behavior stay the same across all of them. Register the chosen size in the ollama provider block in atlas.json with an appropriate "limit": { "context": 131072 } and Atlas will drive whichever variant is present.
Why does a visible reasoning trace help debug Atlas agents?
DeepSeek-R1 emits its reasoning trace, which is genuinely useful when debugging why an Atlas agent made a bad tool call. Because every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, you see the proposed call and, with DeepSeek-R1, the thinking that produced it.
Most model failures inside an agent harness are opaque. The model proposes an edit to the wrong file and you have no idea why. The reasoning trace from DeepSeek-R1 turns that into a readable record: you can see the moment it misread a retrieval result or attached to the wrong symbol. Combined with the fact that Atlas computes a unified diff for every file edit and surfaces it for approval before writing, you get two independent views of the model's intent before anything is written to disk.
How much context does DeepSeek-R1 have on each variant?
DeepSeek-R1 offers 128K to 164K tokens of context depending on the variant, so the window is not a single number the way it is for most models. Set the limit you actually get in atlas.json, commonly "limit": { "context": 131072 }, rather than assuming the largest figure.
The variable window is a direct consequence of the DeepSeek-R1 tag family spanning 1.5B to 671B. Different builds expose different limits, and Atlas will trust the number you configure. Since Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, an over-declared limit means retrieval results get pushed at a model that cannot hold them, and reasoning quality collapses without an obvious error. Declaring the honest number is the cheapest reliability win available with DeepSeek-R1.
Are the small DeepSeek-R1 distills actually usable?
The distilled 7B and 14B variants of DeepSeek-R1 bring reasoning to hardware that cannot host any other reasoning model, which is a real win. The caution is that these small distilled variants keep the reasoning style but lose much of the capability, which can be misleading.
A 7B DeepSeek-R1 will produce a long, confident, well-structured reasoning trace and then reach a wrong conclusion. The format survives the distillation, the judgment does not, and that mismatch is more dangerous than a small model that simply looks weak. If you are running DeepSeek-R1 at 7B or 14B inside Atlas, read the trace critically rather than treating its length as evidence. On a 16GB card, `ollama pull deepseek-r1:14b` is the sensible floor for work you intend to trust.
When should you pick a different model than DeepSeek-R1?
Pick a different model than DeepSeek-R1 for the implementation pass. Long reasoning traces are slow to generate locally, making interactive TUI use painful below a 32B variant on a good GPU, and reasoning is wasted on work where the fix is already understood.
The pattern that works with DeepSeek-R1 inside Atlas is to treat it as a diagnostic instrument, not a daily driver. Select it from /models when you hit a bug a non-reasoning model cannot crack, read the trace, then switch back to qwen3-coder:30b with `model.cycle_recent` for the actual implementation pass. If you would rather not host the reasoning yourself at all, DeepSeek-R1 is served via Bedrock at $1.35 / $5.40 per Mtok, which is expensive next to a coding model but priced for occasional use.
Setup
- 01Pick a size for your hardware: `ollama pull deepseek-r1:14b` for a 16GB card, or `ollama pull deepseek-r1:32b` for 24GB.
- 02Register it in the ollama provider block in atlas.json with an appropriate "limit": { "context": 131072 }.
- 03Select DeepSeek-R1 from /models when you hit a bug a non-reasoning model cannot crack.
- 04Read the emitted reasoning trace to understand why the Atlas agent proposed a given tool call.
- 05Switch back to qwen3-coder:30b with `model.cycle_recent` for the actual implementation pass.
Frequently asked questions
- what sizes does deepseek-r1 come in on ollama
- DeepSeek-R1 is available at 1.5B, 7B, 8B, 14B, 32B, 70B, and 671B under a single `ollama pull deepseek-r1` tag family, so you can right-size to your hardware without changing prompts.
- which deepseek-r1 variant should i run on a 16gb gpu
- Run `ollama pull deepseek-r1:14b` for a 16GB card, or `ollama pull deepseek-r1:32b` if you have 24GB. Below 32B on a good GPU, long reasoning traces are slow enough to make interactive TUI use painful.
- are the small deepseek-r1 distills any good
- The distilled 7B and 14B variants bring reasoning to hardware that cannot host any other reasoning model, but they keep the reasoning style while losing much of the capability, which can be misleading. Read the trace critically.
- how much does deepseek-r1 cost on bedrock
- DeepSeek-R1 is $1.35 / $5.40 per Mtok via Bedrock. Self-hosting through Ollama is free, which is why most Atlas users run it locally and reserve Bedrock for occasional use.
- what is the context window of deepseek-r1
- DeepSeek-R1 has a 128K to 164K token context window depending on the variant. Register the honest number in atlas.json, commonly "limit": { "context": 131072 }, so Atlas does not overfill the model.
- why is my local reasoning model so slow in the terminal
- Long reasoning traces are slow to generate locally. With DeepSeek-R1, interactive TUI use is painful below a 32B variant on a good GPU, because every turn regenerates a full chain of thought.
- how do i switch back to a coding model after debugging
- Use `model.cycle_recent` to switch from DeepSeek-R1 back to qwen3-coder:30b for the implementation pass. Atlas lets you switch the active model and provider on the fly with favorites and recents.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas vs Codebuff: Terminal AI Coding Agents in 2026
Atlas and Codebuff are terminal AI coding agents for 2026. Compare Atlas's terminal-native TUI, permission-gated tools, and diff review with Codebuff's multi-agent system and flexible pricing.
Atlas for Express in 2026
Atlas is a terminal-native AI coding agent for Express in 2026. It reads your middleware order, wraps async route handlers, and runs mocha behind a prompt.
Atlas vs Aider: Terminal AI Coding Agents in 2026
Comparing Atlas and Aider for 2026 developers. Atlas offers a rich TUI, permission-gated tool calls, and plugin support, while Aider provides a repo-map and commits every change as a discrete git revision.
Atlas vs OpenHands: Terminal AI Coding Agents in 2026
Atlas and OpenHands comparison for 2026. Explore terminal-native TUI, self-hosting, code indexing, change review, and pricing models for AI coding agents.
Atlas for R: A Terminal-Native AI Coding Agent for tidyverse, roxygen2, and testthat in 2026
Atlas is a terminal-native AI coding agent for R in 2026. It reads roxygen2 docblocks and renv.lock, rewrites loops as dplyr or purrr pipelines, and runs devtools::test().
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 Erlang in 2026
Atlas is a terminal-native AI coding agent for Erlang/OTP in 2026. Run it in an app with a rebar.config, map supervisors and gen_server modules, review every diff.
Atlas for React Native: Terminal-Native AI Coding Across the Native Boundary in 2026
Atlas is a terminal-native AI coding agent for React Native in 2026. Work across the New Architecture, native modules, and platform-specific files with diff-first review.