Locally Explains

Your local coding AI can edit the wrong files confidently

Every figure and product claim that appears on screen, with the page it comes from. Pages were checked on 28 September 2026.

Ollama can launch coding agents against local models. “ollama launch is a new command which sets up and runs your favorite coding tools like Claude Code, OpenCode, and Codex with local or cloud models.” Commands shown include ollama launch claude and ollama launch opencode. Supported integrations listed: Claude Code, OpenCode, Codex, Droid. https://ollama.com/blog/launch
Launch commands per model. The Qwen3 Coder library page lists applications with commands such as ollama launch claude --model qwen3-coder and ollama launch opencode --model qwen3-coder. https://ollama.com/library/qwen3-coder
Tool calling in Ollama. “Ollama supports tool calling (also known as function calling) which allows a model to invoke tools and incorporate their results into its replies.” https://docs.ollama.com/capabilities/tool-calling
LM Studio. Its documentation describes running models such as Llama, DeepSeek, Qwen and Phi locally, with a chat interface and model search and download (“Use a simple and flexible chat interface”, “Search & download functionality”). https://lmstudio.ai/docs/app

Continue runs in VS Code and JetBrains. “Open source AI code assistant for VS Code and JetBrains.” https://docs.continue.dev/
Model roles. “Models in Continue can be configured to be used for various roles in the extension”: chat, autocomplete, edit, apply, embed (and rerank). https://docs.continue.dev/customize/model-roles/00-intro
Autocomplete recommendation. The autocomplete guide lists QwenCoder2.5 (1.5B) as a best open model for autocomplete and recommends ollama run qwen2.5-coder:1.5b for local use. https://docs.continue.dev/customize/deep-dives/autocomplete
Context length warning. Under “Model requires more system memory to run”: “Continue may use a higher default context length than other tools. Reduce contextLength in your config (e.g., to 2048), or try a smaller model.” https://docs.continue.dev/guides/ollama-guide
Agent mode depends on tool support. Under “Agent mode is not supported”: add capabilities: [tool_use] to the model config, and “If still not working, the model may not actually support tools.” The same guide shows an error reading “does not support tools”. https://docs.continue.dev/guides/ollama-guide
contextLength in config. The same guide’s advanced settings example sets contextLength: 8192, which is the value used in the illustrative config on screen.

OpenCode. “The open source AI coding agent.” Its built in tools include bash (“Execute shell commands”), edit (“Modify existing files”), grep and glob. https://opencode.ai/ and https://opencode.ai/docs/tools/

Qwen3 Coder 30B. qwen3-coder:30b is listed at 19GB with a 256K context window, and the page states it “offers 30B total parameters with only 3.3B activated”. It also lists long context support “optimized for repository-scale understanding” and agentic capabilities for software engineering tasks. https://ollama.com/library/qwen3-coder
Qwen3 Coder’s design. The Qwen team describes it as “our most agentic code model to date”, with results on agentic coding and agentic tool use. https://qwenlm.github.io/blog/qwen3-coder/
Qwen 2.5 Coder sizes. Tags 0.5b (398MB), 1.5b (986MB), 3b (1.9GB), 7b (4.7GB), 14b (9.0GB), 32b (20GB). https://ollama.com/library/qwen2.5-coder
GPT OSS 20B. OpenAI describes gpt-oss-20b as “our medium-sized open-weight model for low latency, local, or specialized use-cases (21B parameters with 3.6B active parameters)”, with configurable reasoning effort. https://developers.openai.com/api/docs/models/gpt-oss-20b Ollama lists gpt-oss:20b at 14GB with a 128K context window and describes the family as designed “for powerful reasoning, agentic tasks, and versatile developer use cases”. https://ollama.com/library/gpt-oss
Devstral Small. “Devstral excels at using tools to explore codebases, editing multiple files and power software engineering agents.” Model size 24B parameters. https://huggingface.co/mistralai/Devstral-Small-2505 Mistral’s announcement and its SWE-Bench Verified chart: https://mistral.ai/news/devstral

A 16GB graphics card. Current consumer cards with 16GB of graphics memory include the GeForce RTX 5070 Ti (“16 GB GDDR7”). https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5070-family/ The 19GB qwen3-coder:30b download is larger than 16GB before context and runtime memory are added. How much memory context, runtime overhead, the operating system and the editor take depends on the setup and is not stated by any of these pages.

No model in this video was run against the invoice bug. The bug, the repository and every file name in it are an illustrative fixture, and the scorecard is shown unscored.
The relative speed of a 1.5B and a 30B model for autocomplete is illustrated, not measured.
The claim that some models advertising tool support still fail at it in practice rests on Continue’s troubleshooting guidance, not on a test of any named model.