Chat privately with Qwen3-Coder-Next in the Ollama app on Windows or Mac, or its interactive Terminal on Linux. Choose official Q4, higher-precision Q8, or an uncensored package, then start a conversation with a practical 4K context.
Estimated starting points for one local chat at 4K context. Smaller GPUs can use system RAM; slower replies are expected with offload. These are capacity estimates, not speed measurements.
| Machine | Qwen3-Coder-Next 80B · Q4 | Qwen3-Coder-Next 80B · Q8 | Qwen3-Coder-Next 80B · Uncensored Q4 |
|---|---|---|---|
| NVIDIA RTX 5090 · 32 GB VRAM + 64 GB RAM | Should run with CPU offload; slower replies | Prefer 128 GB RAM; choose Q4 here | Should run with CPU offload; slower replies |
| NVIDIA GPU · 64 GB VRAM + 64 GB RAM | Recommended capacity at 4K | Prefer 128 GB RAM; offload needed | Recommended capacity at 4K |
| NVIDIA GPU · 96 GB VRAM + 128 GB RAM | Ample capacity at 4K | Recommended capacity at 4K | Ample capacity at 4K |
| AMD GPU · 24 GB VRAM + 64 GB RAM | May run with offload; confirm ROCm/Vulkan | Prefer 128 GB RAM; substantial offload | May run with offload; confirm ROCm/Vulkan |
| Intel Arc · 24 GB VRAM + 64 GB RAM | May run with offload; confirm Vulkan | Prefer 128 GB RAM; substantial offload | May run with offload; confirm Vulkan |
| Apple Silicon · 64 GB unified memory | Tight; close other apps and start at 4K | Not a practical starting configuration | Tight; close other apps and start at 4K |
| Apple Silicon · 96 GB unified memory | Recommended capacity at 4K | Tight; prefer 128 GB unified memory | Recommended capacity at 4K |
| Apple Silicon · 128 GB unified memory | Ample capacity at 4K | Recommended capacity at 4K | Ample capacity at 4K |
| CPU only · 64 GB RAM | Tight; expect slow replies | Prefer 128 GB RAM | Tight; expect slow replies |
| CPU only · 128 GB RAM | Capacity available; expect slow replies | Capacity available; expect slow replies | Capacity available; expect slow replies |
AMD/Intel results depend on a supported driver and Ollama backend. Check ollama ps for actual GPU/CPU split. Package size excludes context, runtime and other applications. More context needs more memory; no exact speed is promised.
Every supplied preset starts at 4,096 tokens. That space covers the conversation and reply; a model's 256K maximum does not mean your machine has room for it.
| Machine | Qwen3-Coder-Next 80B · Q4 | Qwen3-Coder-Next 80B · Q8 | Qwen3-Coder-Next 80B · Uncensored Q4 |
|---|---|---|---|
| GPU · 32 GB VRAM + 64 GB RAM | Start at 4,096 tokens; offload expected | Prefer 128 GB RAM; use Q4 here | Start at 4,096 tokens; offload expected |
| Apple Silicon · 64 GB unified memory | Start at 4,096 tokens; limited headroom | Use Q4; prefer 128 GB for Q8 | Start at 4,096 tokens; limited headroom |
| Q4 · 64 GB VRAM or 96 GB unified memory | Start at 4,096; try 8K after checking memory | Prefer 96 GB VRAM + 128 GB RAM, or 128 GB unified | Start at 4,096; try 8K after checking memory |
| Q8 · 96 GB VRAM + 128 GB RAM, or 128 GB unified | Start at 4,096; try 8K after checking memory | Start at 4,096; try 8K after checking memory | Start at 4,096; try 8K after checking memory |
Start at 4K in the app’s context slider or with /set parameter num_ctx 4096 in Linux interactive chat. Try 8K only with spare memory; reduce it again if replies slow down or allocation fails. These are starting estimates, not benchmark results.
Chat privately with Qwen3-Coder-Next in the Ollama app on Windows or Mac, or its interactive Terminal on Linux. Choose official Q4, higher-precision Q8, or an uncensored package, then start a conversation with a practical 4K context.
Select the weights you want. The instructions, downloads and settings change with your choice.
Save your machine in My Hardware to get an automatic starting choice. You can always choose any package yourself.
52 GB download. Estimated for 4K chat: a 32 GB GPU with 64 GB system RAM can use CPU offload; expect slower replies. About 64 GB VRAM is a full-GPU starting estimate. A 64 GB Mac is tight; 96 GB unified memory is recommended.
Choose where you want to run it. The instructions below update to match your selection.
Use the native app for this exact Qwen3-Coder-Next 80B · Q4 package. Apple Silicon uses Metal; Intel Macs run on the CPU. No separate server or configuration file is required.
Already using the current Ollama app? Open it and skip installation. Otherwise download the Mac app, drag it into Applications and open it on macOS 14 or newer. Allow at least 65 GB free storage for this 52 GB package. A 64 GB Mac is tight; 96 GB is recommended. A 32 GB GPU also needs 64 GB system RAM and CPU offload, which makes replies slower. Keep Ollama open while chatting.
Download Ollama for Mac Installation and supported hardwareIn the app, start a new chat and open the model picker. Search for qwen3-coder-next:q4_K_M and select that exact model, including the tag after the colon. The app can download a registry model when first used; allow the 52 GB download to finish. Do not select a similarly named cloud model or a different quantization. Ollama manages its own model files.
Exact selected package Ollama desktop chatOpen Ollama Settings and set the context-length slider to 4K (4096 tokens) before sending your first message. This is the starting memory allowance used in the tables above. Longer context holds more conversation but uses more memory; try 8K later only if the 4K chat is responsive and your machine has headroom. The model’s maximum context is not a recommended setting for every machine.
Ollama context settingsIn the chat box, type: “Remember that my test word is cedar. What is 7 plus 5?” Send the prompt and check that the answer is 12. In the same chat, send the follow-up: “What test word did I ask you to remember?” Check the reply says cedar. Continue using this conversation for your own questions; this checks that the selected model is answering and following the previous message.
Use the app’s conversation list to reopen this chat, or start a new chat with qwen3-coder-next:q4_K_M for a fresh topic. Wait for any active reply before closing the app. Quit Ollama fully from its menu-bar icon when you want to stop the local service; reopen it from Applications next time. The model stays downloaded. Changing packages should start a new conversation.
Your selected local model answers 12 and then recalls cedar in the same conversation. On Windows/Mac you can reopen that conversation in the app; on Linux continue at the >>> prompt and use /bye when finished. Hardware estimates remain separate from this simple task check.
Every option contains the 80B-total / 3B-active model: 3B active does not make the download or memory requirement that of a 3B model. Download only the package you need. Official and uncensored Q4 each use about 52 GB storage; Q8 uses about 85 GB. Memory ratings are estimates for one 4K conversation.
Update and reopen Ollama, then search the full model tag shown in your selected package. If it still does not appear, open PowerShell on Windows or Terminal on Mac and run ollama pull followed by that exact tag (for example, ollama pull qwen3-coder-next:q4_K_M for official Q4). Return to the app and select the downloaded model. This fallback is only for a failed app download; no separate server or Modelfile is needed.
Choose Q4, lower the app context slider to 4K, and close other loaded models and memory-heavy apps. On Linux set /set parameter num_ctx 4096 inside chat. Q4 still needs roughly 52 GB for its weights; a 32 GB GPU uses system RAM too. Prefer 96 GB Apple unified memory for Q4 and 128 GB for Q8.
On a systemd install, start the service with sudo systemctl start ollama. On a Linux machine without that service, run ollama serve in a second Terminal and leave it open, then retry ollama run in the first. Do not start another server if one is already running.
Check your exact GPU and OS against the linked Ollama GPU support page and install its documented driver/backend. Restart Ollama. For diagnosis only, run ollama ps in PowerShell or Terminal while a chat is loaded to see the GPU/CPU split and context. A 32 GB GPU cannot hold this whole Q4 model; partial offload is expected.
Original instructions, model files, and compatibility notes behind this setup.
Source-backed package and platform-route plan for private local chat with Qwen3.5 · 9B. Preserves the official Ollama 9B package, a smaller 4B family fallback, and a clearly labeled community uncensored GGUF option.
Run Qwen3.8 27B with Ollama, connect it to Cline in VS Code, and make one reviewable code change in a tiny local project.
Generate images from text prompts in ComfyUI with the fast distilled or more adjustable base 4B model. Save your results and reuse the visual workflow.