DeepSeek-V3.1 671B (Ollama) is DeepSeek's hybrid thinking and non-thinking flagship, self-hosted: a 671B mixture-of-experts at 404GB in Q4_K_M with a context window of 160K tokens (163,840). Inside Atlas, one provider entry covers both planning and fast editing because both modes live in one set of weights. Pricing is Free (self-hosted), hardware cost dominates. Only a serious server or a very large unified-memory machine will run it.
What is DeepSeek-V3.1 671B best at inside Atlas?
DeepSeek-V3.1 671B (Ollama) is best at being the only model in an Atlas deployment. Hybrid thinking and non-thinking modes live in one set of 671B weights, so a single provider entry covers both planning and fast editing, at a 160K token (163,840) context.
Most local Atlas setups juggle two tags: a reasoner for the read-only plan agent and a coder for the build agent. DeepSeek-V3.1 671B collapses that. The hybrid design means one registration in the atlas.json ollama models map serves the plan-then-build handoff without a model swap, and the 160K token (163,840) window is large enough that the plan, the retrieved code, and the tool transcript all stay resident across the handoff. Pin model to ollama/deepseek-v3.1:671b and confirm the provider resolved with atlas models ollama. Terminus refresh tags are published alongside the base release.
What hardware does DeepSeek-V3.1 671B actually need?
DeepSeek-V3.1 671B (Ollama) needs roughly 406GB to serve at Q4_K_M, and 713GB at q8_0. That is a datacenter deployment, not a desk. Provision the memory first, then run ollama pull deepseek-v3.1:671b, which is a 404GB download.
The memory figure is the entire story with DeepSeek-V3.1 671B. Roughly 406GB at Q4_K_M rules out every consumer GPU and every ordinary workstation. What remains is a serious multi-GPU server or a very large unified-memory machine. Plan the deployment as infrastructure: serve Ollama on that host, make http://<host>:11434/v1 reachable from the workstation running Atlas, and set the ollama provider options.baseURL in atlas.json to point at it. Atlas itself stays on your laptop; the 404GB of weights do not. If the hardware is not there, the only alternative Ollama offers is a 671b-cloud tag, which is no longer local.
What does DeepSeek-V3.1 671B cost to run?
DeepSeek-V3.1 671B (Ollama) is Free (self-hosted), hardware cost dominates. There is no per-token bill, but roughly 406GB of memory at Q4_K_M, or 713GB at q8_0, is a capital expense. In 2026 that is the honest tradeoff on frontier-class open weights.
Calling DeepSeek-V3.1 671B free is true and misleading at once. The weights cost nothing and the tokens cost nothing, and if you already own the hardware, an unlimited Atlas agent loop on a 671B model is genuinely zero marginal cost. But the entry ticket is a machine that can hold 406GB, and for most teams that is the dominant number in the decision. Do the arithmetic against your actual volume before provisioning. The people for whom this makes sense usually have a non-financial reason as well, which is the next section.
Why self-host DeepSeek-V3.1 671B instead of using a hosted API?
Self-host DeepSeek-V3.1 671B (Ollama) when zero egress is a hard requirement. Open weights you can run behind your own firewall are the only way to get this class of model with no third-party data handling at all, and Atlas keeps the index local too with Ollama embeddings.
The argument for a 406GB deployment is rarely price. It is data handling. DeepSeek-V3.1 671B gives you frontier-class open weights with zero egress, which is a claim a hosted API cannot make regardless of its contract terms. Run the codebase index against Ollama embeddings on the same host so source, vectors, and generations all remain inside your network. Atlas adds the rest of the control surface: every tool call is permission-gated against allow, ask, and deny rules before it runs, and every file edit is surfaced as a unified diff for approval before writing.
When should you pick a smaller model than DeepSeek-V3.1 671B?
Pick a smaller model than DeepSeek-V3.1 671B (Ollama) whenever you do not have roughly 406GB of memory to spare. Q4_K_M needs 406GB and q8_0 needs 713GB, and if the hardware is not there, the only alternative Ollama offers is a 671b-cloud tag, which is no longer local.
DeepSeek-V3.1 671B is the right answer to exactly one question: how do I get frontier open weights with no third-party data handling. For any other question, a smaller local model is a better use of the budget. The 160K token (163,840) window is excellent, but other DeepSeek tags reach the same window on a fraction of the memory. Because Atlas lets you switch the active model and provider on the fly with favorites and recents, you can keep a large remote Ollama host registered for the heavy work and a small local tag for everyday edits, and move between them by provider rather than by rebuild.
Setup
- 01Provision the memory first (roughly 406GB for the Q4_K_M tag), then run: ollama pull deepseek-v3.1:671b (404GB).
- 02Serve Ollama on that host and make http://<host>:11434/v1 reachable from the workstation running Atlas.
- 03In atlas.json set the ollama provider options.baseURL to that host and register deepseek-v3.1:671b with limit.context 163840 and limit.output 16384.
- 04Pin model to ollama/deepseek-v3.1:671b and confirm with atlas models ollama.
- 05Run the codebase index against Ollama embeddings on the same host so source, vectors, and generations all remain inside your network.
Frequently asked questions
- how much memory do i need to run deepseek-v3.1 671b
- Roughly 406GB to serve at Q4_K_M, and 713GB at q8_0, against a 404GB download. That is a datacenter deployment, not a desk.
- how to connect atlas to a remote ollama server
- Serve Ollama on the host, make http://<host>:11434/v1 reachable from the workstation running Atlas, then set the ollama provider options.baseURL in atlas.json to that host and confirm with atlas models ollama.
- what is the context window of deepseek-v3.1 671b
- DeepSeek-V3.1 671B (Ollama) has a 160K token (163,840) context window. Register it in atlas.json with limit.context 163840 and limit.output 16384.
- does deepseek-v3.1 support thinking and non-thinking modes
- Yes. DeepSeek-V3.1 671B ships hybrid thinking and non-thinking modes in one set of weights, so a single Atlas provider entry covers both planning and fast editing.
- is deepseek-v3.1 671b free
- Pricing is Free (self-hosted), hardware cost dominates. There is no per-token bill, but roughly 406GB of memory at Q4_K_M is the entry ticket.
- can i run deepseek-v3.1 671b on a consumer gpu
- No. DeepSeek-V3.1 671B needs roughly 406GB to serve at Q4_K_M. If the hardware is not there, the only alternative Ollama offers is a 671b-cloud tag, which is no longer local.
- best open weight model for zero egress coding agents
- DeepSeek-V3.1 671B (Ollama) gives frontier-class open weights you can run behind your own firewall, with a 160K token (163,840) context. Pair it with a codebase index built on local Ollama embeddings so source, vectors, and generations all remain inside your network.
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.
Atlas vs Devin: AI Coding Agents Compared for 2026
Atlas and Devin offer distinct AI coding experiences in 2026. Atlas provides a terminal-native TUI with local control, while Devin is a cloud-managed engineer with SWE-1.7.
Atlas for F#: A Terminal-Native AI Coding Agent for .fsproj Solutions in 2026
Atlas is a terminal-native AI coding agent for F# in 2026. It respects .fsproj file order, maps discriminated unions, runs dotnet test behind a prompt, and runs Fantomas.
Atlas vs PearAI: Choosing Your AI Coding Agent in 2026
Compare Atlas, the terminal-native AI coding agent, with PearAI, a VS Code fork. Evaluate features, pricing, and community support for developers in 2026.
Atlas for Assembly: Registers, Calling Conventions, and nasm in 2026
Atlas is a terminal-native AI coding agent for Assembly in 2026. It reads .asm and .S sources, tracks System V and AAPCS64 calling conventions, and assembles with nasm behind a prompt.
Write Unit Tests for Untested Code with Atlas in 2026
How to write unit tests for untested code with Atlas in 2026: the lsp tool enumerates exported symbols, grep copies repo conventions, and bash actually runs the suite.
Atlas vs Kiro in 2026: Terminal Agent Compared to AWS's Spec-Driven IDE and CLI
Atlas vs Kiro in 2026. Kiro writes EARS-notation specs before code and charges credits; Atlas is a free, open source terminal agent with diff-before-write review.
Atlas vs GitHub Copilot: Terminal AI Coding Agents in 2026
Comparing Atlas, the terminal-native AI coding agent, with GitHub Copilot's editor extension and chat features for developers in 2026. Explore planning, pricing, and privacy.