Phi-4 (local via Ollama) is Microsoft's 14B model trained heavily on synthetic curriculum data, and inside Atlas it is best used as the cheap fast slot rather than the build agent. Phi-4 is Free (self-hosted), because Ollama serves the weights on your own machine, and it carries a 16K tokens (16,384) context window. That 16,384 token ceiling is the binding constraint: it cannot hold a large Atlas search result plus a file plus a diff, which is exactly what a build turn needs.
What is Phi-4 best at inside Atlas?
Phi-4 (local via Ollama) is best inside Atlas as the small_model slot: titles, summaries, and short reasoning passes. Phi-4 is 14B parameters that match much larger models on math and reasoning benchmarks, a direct result of Microsoft's synthetic-textbook training approach, and it is Free (self-hosted).
Atlas runs more than one model in a session. The build agent does the heavy edit work, while a cheaper model handles session titles, summaries, and the short auxiliary calls that pile up over an hour of work. Phi-4 (local via Ollama) is a strong fit for that second job. At 14B, Phi-4 is fast enough locally that the Atlas TUI stays responsive, which cannot be said for 70B-class local models. Because Phi-4 is Free (self-hosted), moving those auxiliary calls onto it takes the running cost of the non-build half of an Atlas session to zero, while the build agent keeps whatever model you actually trust with a diff.
Why is Phi-4's 16K context window the real constraint?
Phi-4 (local via Ollama) has a 16K tokens (16,384) context window, and that is the binding constraint inside Atlas. A 16,384 token budget cannot hold a large Atlas search result plus a file plus a diff, and an Atlas build turn routinely needs all three at once.
Consider what Atlas actually puts in front of a model on a build turn. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and it indexes code by AST declarations using tree-sitter, so a search returns real declaration bodies rather than trimmed line windows. Add the file being edited, the unified diff Atlas computes for every file edit, and the running conversation, and 16,384 tokens is gone. Register the limit honestly with "limit": { "context": 16384 } in the ollama provider block in atlas.json so Atlas plans around the real ceiling. Then use Phi-4 where 16K is genuinely enough: summarizing, naming, classifying, and short single-file reasoning.
How much memory does Phi-4 need to run Atlas locally?
Phi-4 (local via Ollama) is roughly 9GB at 4-bit, so it fits comfortably on a 12GB GPU alongside an embedding model for Atlas's local codebase index. That 9GB footprint is the practical reason Phi-4 shows up on machines that cannot host a 70B model at all.
The 9GB at 4-bit figure matters because Atlas wants two models resident at once on a local setup. Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers, which means an embedding model needs GPU memory too. On a 12GB GPU, Phi-4 at roughly 9GB leaves headroom for that embedding model, so the hybrid semantic and keyword retrieval keeps working without swapping models in and out. Phi-4 is Free (self-hosted), so the whole arrangement, chat model and embeddings, has no per-token cost at all, and nothing leaves the machine.
Is Phi-4 actually good at real-world code?
Phi-4 (local via Ollama) is benchmark-strong but weaker on messy real-world code than its scores suggest, which is a known criticism of synthetic-data training. The 14B model matches much larger models on math and reasoning benchmarks, and then underdelivers on the kind of code Atlas is pointed at.
Be blunt about this before wiring Phi-4 (local via Ollama) into a build agent. The Phi line's whole thesis is synthetic curriculum data, and it produces genuinely strong reasoning benchmark scores at 14B. Real repositories are not curriculum. They have dead code, inconsistent naming, and half-migrated abstractions, and Phi-4 handles that less well than its benchmark numbers imply. Atlas mitigates some of the risk structurally: Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs. But approval gates catch bad edits, they do not make a model produce good ones.
When should you pick a different model instead of Phi-4?
Pick a different model than Phi-4 (local via Ollama) whenever an Atlas turn needs more than 16,384 tokens, which is most build turns. Phi-4's 16K window cannot hold a large Atlas search result plus a file plus a diff, so reserve Phi-4 for the small_model slot instead.
The clean split is to keep Phi-4 (local via Ollama) as small_model for titles and summaries rather than as the build agent, exactly as the Atlas setup notes recommend, and give the build agent a model with real context headroom. Atlas lets you switch the active model and provider on the fly with favorites and recents, so this is a configuration decision, not a fork in the road: register phi4 once, star it, and flip the build agent in `/models` when a task grows past what 16,384 tokens can carry. The value of Phi-4 is that its half of the session stays free and fast.
Setup
- 01Pull the weights with `ollama pull phi4` (roughly 9GB at 4-bit).
- 02Register it under the ollama provider block in atlas.json with "limit": { "context": 16384 }.
- 03Select it in `/models` inside the Atlas TUI.
- 04Given the 16K window, use it as `small_model` for titles and summaries rather than as the build agent.
Frequently asked questions
- how to set up phi-4 with atlas
- Run `ollama pull phi4`, register it under the ollama provider block in atlas.json with "limit": { "context": 16384 }, then select it in `/models`. Given the 16K window, set it as `small_model` rather than the build agent.
- what is phi-4's context window
- Phi-4 has a 16K tokens (16,384) context window. Inside Atlas that is the binding constraint, because 16,384 tokens cannot hold a large Atlas search result plus a file plus a diff at the same time.
- how much does phi-4 cost
- Phi-4 (local via Ollama) is Free (self-hosted). Ollama serves the 14B weights on your machine, so Atlas pays no per-token cost. The expense is the GPU memory, roughly 9GB at 4-bit.
- can phi-4 run on a 12gb gpu
- Yes. Phi-4 is roughly 9GB at 4-bit, so it fits comfortably on a 12GB GPU alongside an embedding model for Atlas's local codebase index, which Atlas can build with local Ollama embeddings.
- is phi-4 good enough to be an ai coding agent
- Phi-4 is a poor build agent for Atlas because of the 16,384 token ceiling, and it is weaker on messy real-world code than its benchmark scores suggest. Use it as `small_model` for titles and summaries instead.
- why is phi-4 so good at benchmarks but not at coding
- Phi-4 is 14B parameters trained heavily on synthetic curriculum data, which produces reasoning benchmark scores matching much larger models. Real repositories are messier than curriculum data, and that gap is a known criticism of the approach.
- what should i use phi-4 for in atlas
- Use Phi-4 (local via Ollama) as the `small_model` slot in Atlas: session titles, summaries, and short reasoning passes. It is free, fast enough that the Atlas TUI stays responsive, and its 16K window is sufficient for those jobs.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas vs Warp: Choosing Your AI Coding Agent in 2026
Compare Atlas, the terminal-native AI coding agent, with Warp, a smart terminal with AI Agent Mode, for developers in 2026. Evaluate features, privacy, and workflow.
Atlas for Axum in 2026
Atlas is a terminal-native AI coding agent for Axum in 2026. It decodes tower trait-bound errors, adds IntoResponse types, and runs cargo nextest run.
Atlas vs Tabnine: Choosing Your Terminal AI Coding Agent in 2026
Comparing Atlas and Tabnine for developers in 2026. Atlas offers a terminal-native AI agent with diff review, while Tabnine provides privacy-first code completion and chat.
Atlas for FastAPI in 2026
Atlas is a terminal-native AI coding agent for FastAPI in 2026. It reads Pydantic v2 models and Depends(), then runs pytest behind a permission prompt.
Atlas vs Jules: Terminal AI Coding Agents in 2026
Atlas and Jules comparison for 2026. Atlas offers terminal-native TUI and local code indexing. Jules provides cloud VMs and a Critic agent for PR review.
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 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
Atlas for Kotlin in 2026
In 2026, Atlas empowers Kotlin developers with terminal-native AI coding. It integrates with Gradle and coroutines, offering secure, privacy-focused code assistance with local embeddings and granular control.