Phi-3 Medium 14B (Ollama) is Microsoft's Phi-3 Medium, 7.9GB with a 128K tokens (131,072) context on the default tag, priced Free (self-hosted). Phi's whole thesis is data quality over parameter count, and at 14B it punches above its size on reasoning while staying comfortably inside a 12GB card. Inside Atlas that means long tool transcripts fit without a 20GB model. The tag matters: :medium-4k is the same 7.9GB but caps at 4K, so pull phi3:14b or :medium-128k for the long window.
Why does Phi-3 Medium 14B punch above its parameter count?
Phi-3 Medium 14B (Ollama) is trained on heavily curated synthetic and filtered data, which is why it outperforms its parameter count on reasoning tasks. Microsoft's Phi thesis is data quality over parameter count, and at 14B that shows up as reasoning strength from a 7.9GB download.
For Atlas the relevant reasoning work is not trivia, it is deciding what to do next given a tool transcript. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and a model that reasons well for its size is valuable precisely in that plan seat. Phi-3 Medium 14B holds 128K tokens (131,072) while serving in roughly 10GB, a comfortable fit for a 12GB GPU with room for cache. The curated-data approach has a cost, and it is worth stating plainly: narrower world knowledge than a Llama or Qwen of the same size. Phi-3 Medium reasons well about what you show it and knows less about what you do not.
Which Phi-3 tag gives you the 128K window?
Pull phi3:14b (7.9GB, also tagged :medium and :medium-128k) for Phi-3 Medium 14B's full 128K tokens (131,072) window. Avoid :medium-4k unless you want the 4K variant, because it is the same 7.9GB download but caps the context at 4K, which is useless for Atlas agent loops.
The tag trap is the single most common Phi-3 Medium 14B mistake, and it is easy to fall into because the file size gives no warning: :medium-4k and :14b weigh the same. Add phi3:14b to the atlas.json ollama models map with limit.context 131072 and limit.output 8192 so Atlas sizes the prompt for the real window. Set model to ollama/phi3:14b and check resolution with atlas models ollama before you trust a long session to it. If the resolved context comes back small, you pulled the 4K tag.
What can Phi-3 Medium 14B do with 128K tokens in Atlas?
With 128K tokens (131,072) at only 7.9GB, Phi-3 Medium 14B (Ollama) lets long Atlas tool transcripts fit without a 20GB model. That is the practical headline: a full plan-and-build session with many tool calls stays in context on a 12GB card rather than compacting away.
Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and every one of those accumulates output that must go back into the prompt. Phi-3 Medium 14B absorbs that. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, so the code that joins the transcript in that 128K window is ranked rather than arbitrary, and Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so each retrieved chunk is a complete declaration. Roughly 10GB to serve leaves room for cache growth on a 12GB GPU, which is what makes the advertised window usable rather than aspirational.
What does Phi-3 Medium 14B not know?
Phi-3 Medium 14B (Ollama) has narrower world knowledge than a Llama or Qwen of the same size, because its curated-data training trades breadth for reasoning quality. Inside Atlas that means Microsoft's Phi-3 Medium is stronger at reasoning over what you give it than at recalling an obscure library from memory.
The mitigation is retrieval, and Atlas is built for it. Atlas connects to Model Context Protocol servers and exposes their tools to the agent, so a documentation server can supply the facts Phi-3 Medium 14B never memorized. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, so the codebase itself is the knowledge source rather than the weights. Configured this way, the narrow world knowledge of Phi-3 Medium 14B stops being a defect and becomes a shape you design around: put the facts in the 128K tokens (131,072) window and let the model reason.
When should you pick a different model than Phi-3 Medium 14B?
Pick something else when broad recall matters more than reasoning, because Phi-3 Medium 14B (Ollama) has narrower world knowledge than a Llama or Qwen of the same size. Keep phi3:14b when you want 128K tokens (131,072) at 7.9GB and roughly 10GB to serve on a 12GB card.
Phi-3 Medium 14B is a strong fit for Atlas plan agents, for reasoning-heavy review, and for anyone whose GPU is 12GB and whose sessions are long. It is a weaker fit when you need encyclopedic familiarity with a long tail of frameworks. Atlas lets you switch the active model and provider on the fly with favorites and recents, so pairing phi3:14b with a code specialist is a keystroke rather than a rewrite. Everything here is Free (self-hosted), and enabling Ollama embeddings for the codebase index means both the model and its retrieval layer run without a network dependency.
Setup
- 01Pull the long-context tag explicitly: ollama pull phi3:14b (7.9GB, also tagged :medium and :medium-128k). Avoid :medium-4k unless you want the 4K variant.
- 02Add phi3:14b to the atlas.json ollama models map with limit.context 131072 and limit.output 8192.
- 03Expect roughly 10GB of VRAM or unified memory at moderate context.
- 04Set model to ollama/phi3:14b and check resolution with atlas models ollama.
- 05Enable Ollama embeddings for the codebase index so both the model and its retrieval layer run without a network dependency.
Frequently asked questions
- how to run phi-3 medium 14b in atlas
- Run ollama pull phi3:14b (7.9GB), add phi3:14b to the atlas.json ollama models map with limit.context 131072 and limit.output 8192, set model to ollama/phi3:14b, and check resolution with atlas models ollama.
- why is my phi3 context only 4k
- You likely pulled :medium-4k, which is the same 7.9GB download as Phi-3 Medium 14B but caps at 4K. Use phi3:14b or :medium-128k for the 128K tokens (131,072) window.
- what is phi-3 medium 14b's context window
- Phi-3 Medium 14B (Ollama) has a 128K tokens (131,072) context on the default tag, from a 7.9GB download.
- how much vram does phi3:14b need
- Phi-3 Medium 14B takes roughly 10GB of VRAM or unified memory at moderate context, a comfortable fit for a 12GB GPU with room for cache.
- is phi-3 medium better than larger models
- Phi-3 Medium 14B is trained on heavily curated synthetic and filtered data, which is why it outperforms its parameter count on reasoning tasks. The tradeoff is narrower world knowledge than a Llama or Qwen of the same size.
- how much does phi-3 medium 14b cost
- Phi-3 Medium 14B (Ollama) is Free (self-hosted). The only cost is hardware: a 7.9GB download and roughly 10GB to serve.
- can phi-3 medium 14b run atlas fully offline
- Yes. Serve phi3:14b through Ollama and enable Ollama embeddings for the codebase index so both the model and its retrieval layer run without a network dependency.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Flutter in 2026
Discover Atlas for Flutter in 2026. This terminal-native AI coding agent helps Flutter developers build apps faster and safer, integrating with widgets, state, and the Dart toolchain.
Atlas for Spring in 2026
Atlas, the terminal native AI coding agent, empowers Spring developers in 2026 with intelligent code assistance, secure local embeddings, and transparent review processes for enhanced productivity.
Atlas for Clojure: A Terminal-Native AI Coding Agent for deps.edn and Kaocha in 2026
Atlas is a terminal-native AI coding agent for Clojure in 2026. It reads deps.edn aliases and namespace requires, runs clojure -M:test with Kaocha, and applies cljfmt.
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 vs Sourcery: Terminal AI Coding Agents in 2026
Compare Atlas and Sourcery for 2026. Atlas offers terminal-native AI coding with permission-gated tools and local embeddings. Sourcery provides Python-first LLM review and rules-based refactoring, starting at
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 Amazon Q Developer: Terminal AI Coding Agents in 2026
Comparing Atlas and Amazon Q Developer in 2026. Atlas offers terminal-native AI with permission-gated tools and local code indexing. Amazon Q Developer excels in AWS services, providing Java upgrades and security
Atlas vs GitHub Copilot CLI: A Developer's Guide for 2026
Comparing Atlas and GitHub Copilot CLI for developers in 2026. Explore pricing, privacy, agentic workflows, and extensibility to choose the best terminal AI coding agent.