# Atlas with Qwen3-Coder 480B (Ollama): the self-hosted ceiling in 2026 > Qwen3-Coder 480B (Ollama) is 290GB of weights carrying 256K tokens (262,144) of context, Free (self-hosted), but the hardware is not. Qwen3-Coder 480B (Ollama) is the full-size Qwen3-Coder. At 290GB of weights this is a server-class deployment, not a laptop one: think a multi-GPU node or a large unified-memory box. Context is 256K tokens (262,144) and the model is Free (self-hosted), but the hardware is not. Qwen3-Coder 480B is worth documenting because it is the ceiling of what a self-hosted Atlas can run with zero data egress. ## Key takeaways - 290GB Q4_K_M weights carrying the full 256K tokens (262,144) Qwen3-Coder context, the strongest self-hostable coder in the family. - Roughly 292GB to serve: a multi-GPU node or a large unified-memory box, not consumer hardware. - Free (self-hosted), but the hardware is not. - Self-hosting it means an entire agentic coding stack with no vendor in the loop, which is the point for regulated or air-gapped teams. - Same tag family as qwen3-coder:30b, so a prompt and config that work on the 30B port straight up to the 480B. - If you cannot supply the memory, Ollama publishes a 480b-cloud tag, which defeats the local-first premise. ## What hardware does Qwen3-Coder 480B (Ollama) require? Qwen3-Coder 480B (Ollama) is roughly 292GB to serve against 290GB of weights. Verify you actually have the memory across GPUs or unified memory before pulling, because this is a server-class deployment: a multi-GPU node or a large unified-memory box, not consumer hardware. Ollama's own page effectively treats qwen3-coder:480b as a datacenter tag, and the numbers explain why. The pull alone, `ollama pull qwen3-coder:480b`, moves 290GB. Serving takes roughly 292GB across your accelerators or unified memory. There is no quantization trick documented here that brings that under a consumer budget. If you cannot supply that memory, Ollama publishes a 480b-cloud tag instead, which defeats the local-first premise that made you look at a 480B self-hosted model in the first place. ## Why run Qwen3-Coder 480B locally instead of calling a hosted API? Self-hosting Qwen3-Coder 480B (Ollama) means an entire agentic coding stack with no vendor in the loop, which is the whole point for regulated or air-gapped teams. Atlas plus a 290GB local model plus a local embedder produces zero data egress, at 256K tokens (262,144) of context. For a team that cannot send source code to a third party, capability and compliance usually pull in opposite directions. Qwen3-Coder 480B (Ollama) is where they stop fighting: it is the strongest self-hostable coder in the family, so choosing compliance no longer means choosing a weak model. Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers, so run the codebase index with Ollama embeddings on the same host and source and vectors stay inside your own network boundary along with the weights. ## How wide is the Qwen3-Coder 480B context window in Atlas? Qwen3-Coder 480B (Ollama) carries the full 256K tokens (262,144) Qwen3-Coder context on its 290GB Q4_K_M weights. Register qwen3-coder:480b in atlas.json with limit.context 262144 and limit.output 16384, which is the widest window in any locally served Atlas configuration. A 256K window changes what Atlas can hold in a single turn. 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, so the retrieved chunks are dense. With 256K tokens (262,144) available, a large slice of a real repository plus the plan transcript plus the diffs can coexist in one context. Set limit.output to 16384 so a long multi-file diff is not truncated on the way out. ## How do you point Atlas at a remote Ollama host running Qwen3-Coder 480B? Serve Ollama on the host with the 292GB of memory and expose http://:11434/v1 to the workstation running Atlas. In atlas.json point the ollama provider options.baseURL at that host, register qwen3-coder:480b, and set model to ollama/qwen3-coder:480b. Nobody runs Qwen3-Coder 480B (Ollama) on the same machine as their editor. The split is the normal deployment: a server-class box holds the 290GB of weights, and Atlas runs on your workstation talking to it over http://:11434/v1. Check resolution with `atlas models ollama` after you wire it up. Keep the codebase index on the same host with Ollama embeddings so source and vectors stay inside your own network boundary rather than crossing it in one direction and not the other. ## When is Qwen3-Coder 480B the wrong choice? Qwen3-Coder 480B (Ollama) is the wrong choice for anyone without roughly 292GB of memory across GPUs or unified memory. Ollama publishes a 480b-cloud tag for that case, but the cloud tag defeats the local-first premise, so a smaller local Qwen3-Coder size is usually the better answer. The saving grace is portability inside the family. Qwen3-Coder 480B (Ollama) shares a tag family with qwen3-coder:30b, so a prompt and config that work on the 30B port straight up to the 480B. That means a team can develop an Atlas configuration on affordable hardware and move it to a 480B node later without rewriting anything. Start where your hardware is. The 480B is the ceiling to grow into, not the entry point, and Free (self-hosted) never meant free to run. ## Setup 1. Verify you actually have the memory (roughly 292GB across GPUs or unified memory) before pulling. 2. Run: ollama pull qwen3-coder:480b (290GB). 3. Serve Ollama on the host and expose http://:11434/v1 to the workstation running Atlas. 4. In atlas.json point the ollama provider options.baseURL at that host and register qwen3-coder:480b with limit.context 262144 and limit.output 16384. 5. Set model to ollama/qwen3-coder:480b and check resolution with atlas models ollama. 6. Run the codebase index with Ollama embeddings on the same host, so source and vectors stay inside your own network boundary. ## FAQ ### how much memory does qwen3-coder 480b need Roughly 292GB to serve Qwen3-Coder 480B (Ollama) against 290GB of weights. Verify the memory across GPUs or unified memory before pulling, because this is a server-class deployment, not consumer hardware. ### can i run qwen3-coder 480b on a consumer gpu No. Qwen3-Coder 480B (Ollama) is 290GB of weights and roughly 292GB to serve. Ollama's own page effectively treats it as a datacenter tag. If you cannot supply that memory, Ollama publishes a 480b-cloud tag, which defeats the local-first premise. ### what context window does qwen3-coder 480b have Qwen3-Coder 480B (Ollama) carries the full 256K tokens (262,144) Qwen3-Coder context. Register it in atlas.json with limit.context 262144 and limit.output 16384 so long multi-file diffs are not truncated. ### how do i connect atlas to a remote ollama server Serve Ollama on the host and expose http://:11434/v1 to the workstation running Atlas. In atlas.json point the ollama provider options.baseURL at that host, register qwen3-coder:480b, then check resolution with atlas models ollama. ### is qwen3-coder 480b free Qwen3-Coder 480B (Ollama) is Free (self-hosted), but the hardware is not. The weights cost nothing; the roughly 292GB of GPU or unified memory needed to serve them is the real bill. ### can i develop on qwen3-coder 30b and move to 480b later Yes. Qwen3-Coder 480B (Ollama) is the same tag family as qwen3-coder:30b, so a prompt and config that work on the 30B port straight up to the 480B. Develop on affordable hardware and move to the 480B node when it exists. ### what is the best local model for an air-gapped coding agent Qwen3-Coder 480B (Ollama) is the ceiling of what a self-hosted Atlas can run with zero data egress. Self-hosting it means an entire agentic coding stack with no vendor in the loop, which is the whole point for regulated or air-gapped teams. --- Canonical HTML: https://runatlas.sh/resources/models/qwen3-coder-480b-local Source of truth: aeo_pages row `/resources/models/qwen3-coder-480b-local` (segment: Models) (this file is generated from it, never hand-edited). Licence: Atlas is proprietary with a free core. It is not open source and there is no public source repository.