Qwen3 30B-A3B (Ollama) is a mixture-of-experts model with 30B total parameters but only about 3B active per token. You pay 19GB of memory for the weights and get inference speed closer to a small model, which is the single best throughput trade available for a local Atlas agent. Context is 256K tokens (262,144) on the instruct and thinking tags, and pricing is Free (self-hosted).
What does 30B-A3B mean for Atlas inference speed?
Qwen3 30B-A3B (Ollama) carries 30B total parameters with roughly 3B active per token. You pay the memory cost of a 30B and get the token rate of something far smaller, which is the single best throughput trade available for a local Atlas agent in 2026.
Mixture-of-experts routing means most of the 30B parameters sit idle on any given token while a small subset fires. For Atlas, where a single task can mean dozens of tool calls and re-reads, tokens per second is the difference between an agent you use and an agent you wait on. Qwen3 30B-A3B (Ollama) gives you 30B-class knowledge at roughly 3B-class latency. Every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, so a fast model also means less time staring at a spinner between approval prompts.
How much memory does Qwen3 30B-A3B actually need?
Qwen3 30B-A3B (Ollama) needs roughly 21GB of VRAM or unified memory to serve against 19GB of weights. The MoE still needs all 19GB resident, because all experts must be resident even though few fire per token. Small active-parameter count does not mean small memory.
The most common misunderstanding about Qwen3 30B-A3B (Ollama) is that roughly 3B active parameters implies a 3B memory footprint. Routing decides which experts compute, not which experts are loaded. Confirm you have roughly 21GB of VRAM or unified memory free before you pull, then run `ollama pull qwen3:30b-a3b` (19GB). The instruct and thinking tags are the same size, so switching reasoning behavior does not change the memory math. What you are buying with 21GB is speed, not a memory discount.
Does Qwen3 30B-A3B really have a 256K context in Ollama?
Qwen3 30B-A3B (Ollama) advertises 256K tokens (262,144) on the 30b-a3b instruct and thinking tags, enough to hold a serious slice of a repo. Ollama shows 40K rather than 256K on the plain 30b-a3b-q4_K_M tag, so check the tag you actually pulled before setting limit.context.
That discrepancy is a real trap. Registering qwen3:30b-a3b in the atlas.json ollama models map with limit.context 262144 is correct for the instruct and thinking tags, and wrong for the plain 30b-a3b-q4_K_M tag, which Ollama shows at 40K. Atlas will happily send a prompt against whatever limit you declare, so a mismatch produces truncation rather than an error. Check the tag first. When the 256K window is genuinely available, it is enough to hold a serious slice of a repo, which pairs well with the fact that Atlas indexes code by AST declarations using tree-sitter, not blind line windows.
Should you run the instruct or the thinking tag of Qwen3 30B-A3B?
Qwen3 30B-A3B (Ollama) ships instruct and thinking variants at the same 19GB, so reasoning can be toggled by swapping the tag rather than the hardware. Pull both, register both, and use `/models` in the Atlas TUI to add them to favorites for fast switching.
Because the instruct and thinking tags of Qwen3 30B-A3B (Ollama) are identical in size, choosing between them is a behavior decision with no hardware consequence. The thinking tag deliberates before emitting a tool call, which helps when Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent. The instruct tag is faster per turn. Atlas lets you switch the active model and provider on the fly with favorites and recents, so set model to ollama/qwen3:30b-a3b, run `/models` to favorite both tags, and flip based on the task in front of you.
When should you pick a dense model over Qwen3 30B-A3B?
Pick a dense model over Qwen3 30B-A3B (Ollama) when correctness on a single hard problem matters more than tokens per second. Qwen3 30B-A3B spends roughly 3B active parameters per token, which is what makes it fast and also what caps its depth on the hardest one-shot reasoning.
Qwen3 30B-A3B (Ollama) is the right default for interactive Atlas work: long sessions, many tool calls, fast turnaround, and 256K tokens (262,144) of window on the right tags. It is the wrong default for a single gnarly problem where you would happily wait. The memory bill is roughly the same either way at this tier, so the choice is purely about how you want the compute spent. Index the repo with the Ollama embedder so AST chunks, vectors, and prompts all stay local regardless of which tag you settle on.
Setup
- 01Pull the MoE: ollama pull qwen3:30b-a3b (19GB). The instruct and thinking tags are the same size.
- 02Confirm you have roughly 21GB of VRAM or unified memory free; all experts must be resident even though few fire per token.
- 03Check which tag you pulled: Ollama shows 40K rather than 256K on the plain 30b-a3b-q4_K_M tag.
- 04Register qwen3:30b-a3b in the atlas.json ollama models map with limit.context 262144 and limit.output 8192.
- 05Set model to ollama/qwen3:30b-a3b and run /models in the TUI to add it to favorites for fast switching.
- 06Index the repo with the Ollama embedder so AST chunks, vectors, and prompts all stay local.
Frequently asked questions
- what does a3b mean in qwen3 30b-a3b
- A3B refers to roughly 3B active parameters per token. Qwen3 30B-A3B (Ollama) is a mixture-of-experts model with 30B total parameters, so you pay 19GB of memory for the weights and get inference speed closer to a small model.
- how much vram does qwen3 30b-a3b need
- Roughly 21GB of VRAM or unified memory. Qwen3 30B-A3B (Ollama) is 19GB of weights and all experts must be resident even though few fire per token, so the small active-parameter count does not reduce the memory bill.
- why does ollama show 40k context for qwen3 30b-a3b
- Because Ollama shows 40K rather than 256K on the plain 30b-a3b-q4_K_M tag. The 256K tokens (262,144) window is on the 30b-a3b instruct and thinking tags, so check the tag you actually pulled before setting limit.context in atlas.json.
- is a moe model faster than a dense model for coding agents
- Qwen3 30B-A3B (Ollama) fires roughly 3B active parameters per token against 30B total, which gives it the token rate of something far smaller. For Atlas sessions with many tool calls, that is the single best throughput trade available locally.
- qwen3 30b-a3b instruct or thinking tag
- Both ship at the same 19GB, so reasoning can be toggled by swapping the tag rather than the hardware. Register both for Qwen3 30B-A3B (Ollama) and use /models in the Atlas TUI to add them to favorites for fast switching.
- how do i set qwen3 30b-a3b as my atlas model
- Register qwen3:30b-a3b in the atlas.json ollama models map with limit.context 262144 and limit.output 8192, then set model to ollama/qwen3:30b-a3b. Run /models in the TUI to favorite it for fast switching.
- does qwen3 30b-a3b cost anything
- Qwen3 30B-A3B (Ollama) is Free (self-hosted). The cost is hardware: 19GB of weights, roughly 21GB of VRAM or unified memory to serve, and electricity. There is no per-token charge on Atlas sessions.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Actix Web in 2026
Atlas is a terminal-native AI coding agent for Actix Web in 2026. It reads extractors and app_data, then runs cargo test and cargo clippy behind a prompt.
Atlas vs Pieces for Developers: AI Tools for Developers in 2026
Comparing Atlas, a terminal-native AI coding agent, with Pieces for Developers, an OS-level memory layer, for developers in 2026. Evaluate code generation, safety, and context management.
Atlas for WebAssembly: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for WebAssembly. Audit wasm-bindgen exports, batch calls across the JS boundary, and run wasm-pack test --node in 2026.
Rename a Symbol Across the Repo with Atlas in 2026
How to rename a symbol across a repo with Atlas in 2026: findReferences gets the true reference set, grep catches strings and docs, and edit refuses ambiguous matches.
Atlas for React in 2026
Adopt Atlas, the terminal-native AI coding agent, for React development in 2026. Enhance your workflow with intelligent code search, refactoring, and testing for React components and hooks.
Research a Third-Party API Before Integrating It with Atlas in 2026
How to research a third-party API with Atlas in 2026: websearch finds the current docs, webfetch pulls the page as markdown or text, and grep checks repo conventions.
Atlas for OCaml: A Terminal-Native AI Coding Agent for dune and opam Projects in 2026
Atlas is a terminal-native AI coding agent for OCaml in 2026. It reads dune stanzas and .mli signatures, runs dune runtest behind a prompt, and finishes with ocamlformat.
Atlas for Nuxt: Auto-Imports, useAsyncData, and Nitro Handlers in 2026
Atlas is a terminal-native AI coding agent for Nuxt in 2026. It reads nuxt.config.ts, pages/ routes, composables/ auto-imports, and server/api/ Nitro handlers, and tests with @nuxt/test-utils.