Continue.dev with Local LLMs After the Cursor Acquisition: Still Worth It?

By LLM Hardware team · Methodology ·

Key facts

Acquisition
continue.dev says Continue has joined Cursor.
Repo flag
Code is described as still available.
Chat model
27–32B Q4 wants 24 GB. 7B autocomplete fits 8 GB.
Checked
2026-10-13

Checked 2026-10-13. Author: LLM Hardware team. Amazon prices are not stored or shown on this site.

Citeable facts

Key fact: Continue’s site says Continue has joined Cursor, and that the open-source codebase remains freely available.

How we got it: continue.dev, read 2 October 2026. A GitHub “read-only” badge was not confirmed in that pass.

Checked:

Key fact: The extension can still target a local model. Hardware tiers match any other IDE assistant: 7B on 8 GB, 27–32B on 24 GB.

How we got it: Same VRAM formula as the Aider guide. docs.continue.dev documents Ollama and LM Studio providers; re-check those pages if a menu name moved.

Checked:

Key fact: This is not the long-term bet if you are choosing a platform to build on. It is a working local copilot you can still install.

How we got it: Editorial judgment from the acquisition notice, not a deprecation date. No deprecation date was on the homepage on 2 October 2026.

Checked:

Key fact: A 3B autocomplete model and Qwen3-32B as the chat model, loaded together in one Ollama server, need about 23.0 GB at Q4.

How we got it: 3 x 0.6 + 32 x 0.6 + 2, with the 2 GB OS and engine reserve counted once (methodology). Keep-alive and loaded-model defaults per the Ollama FAQ.

Checked:

Methodology explains the VRAM formula, the bytes per weight at each quantization, and the full source list.

Continue is the open-source IDE copilot for VS Code and JetBrains. On 2 October 2026 the homepage said Continue has joined Cursor, thanked the community, and said the open-source codebase remains freely available. The GitHub project is continuedev/continue. This guide does not claim the repository is archived or read-only; that flag was not visible in the 2 October 2026 fetch. The company that was going to keep shipping a rival IDE agent has joined the company that ships Cursor. Treat new feature work as uncertain.

When it is still worth installing

Install it if you want autocomplete and chat inside the editor, pointed at a model on your machine, and you already live in VS Code or JetBrains. The provider docs to re-read are the Ollama guide and the LM Studio provider. If those URLs move, the intent is the same: Continue is a client. Ollama or LM Studio is the server. The GPU requirement does not change because Cursor bought the project.

Skip it as a foundation if you are about to build a team workflow on top of Continue’s config format. Aider and OpenHands are the local tools to start from in that case. Cursor itself, including Cloud Agents, is the commercial product the acquisition points at. Cloud Agents do not use your GPU.

The same memory tiers as any IDE assistant

Tab-complete wants low latency more than it wants a 70B model. A 7B that stays resident will feel better than a 32B that spills to RAM and stalls the keystroke. Chat-with-the-repo wants the larger model and a short file set. You can configure both, and you should not load both at once on 24 GB.

Job in the IDE Model Q4 formula
Inline complete K2-Horizon-7B 6.2 GB. An 8 GB GPU is enough for the weights.
Chat over a few files Qwen3.8-27B or K2-32B 18.2 GB or 21.2 GB. 24 GB card.
Both at once 7B + 32B in one server: 4.2 + 19.2 + 2 = 25.4 GB Over a 24 GB card. A 32 GB GPU or a 36 GB Mac leaves about 6.6 GB for the chat cache, or use a 3B autocomplete model (worked example below). The calculator is one model at a time.

This page does not quote tokens per second for Continue. Once the weights fit, bandwidth sets speed, whichever editor extension sends the request.

Worked example: autocomplete and chat loaded together

Continue can use one model for tab-complete and another for chat. What matters is the cost of keeping both loaded, because an autocomplete model that has to reload before it answers is not much use. Ollama's FAQ says a model stays in memory for 5 minutes after its last request by default, that a negative keep_alive (or OLLAMA_KEEP_ALIVE) keeps it loaded, and that by default up to three models per GPU can be loaded at once (OLLAMA_MAX_LOADED_MODELS).

Count the formula's +2 GB once: it is the operating system and the inference engine, and both models sit in the same Ollama server. A 3B-class autocomplete model at Q4 adds 3 x 0.6 = 1.8 GB.

Both models in one Ollama server, Q4, our VRAM formula

Pair Both loaded What is left for the chat cache
3B + Qwen3-32B 1.8 + 19.2 + 2 = 23.0 GB 24 GB GPU: 1.0 GB, less than one 8K chat cache (2.15 GB). RTX 5090 32 GB or Mac Studio M5 Max 36 GB (about 32 GB usable): 9.0 GB, enough for a 32K f16 cache (8.6 GB).
3B + Qwen3-30B-A3B 1.8 + 18.3 + 2 = 22.1 GB 24 GB GPU: 1.9 GB, about 19,300 tokens of f16 cache. Mac mini M5 Pro 24 GB (about 20 GB usable): does not fit.
3B + Llama 3 8B 1.8 + 4.8 + 2 = 8.6 GB 12 GB GPU or Mac mini M6 16 GB (about 12 GB usable): 3.4 GB, about three 8K chat caches. 16 GB GPU: 7.4 GB.

Adding the two formula results instead (3.8 + 21.2 = 25.0 GB) counts the reserve twice. That figure is closer if you run two separate servers, for example LM Studio for one model and Ollama for the other.

If both stay resident, give the autocomplete model a negative keep_alive so it does not unload after five idle minutes, and leave the chat model on the default if memory is tight: it unloads when idle and loads again on your next question.

Hardware, if you are staying local

A 24 GB card is the interesting purchase, not because Continue requires it, but because the chat model does. RTX 4090, catalog ASIN B0BG94PS2F. A Mac mini M5 Pro 24 GB (B0HGGHNQY6) covers the same fit with less noise. If you only want 7B autocomplete, do not buy either. An 8 GB GPU you already own is the whole story.

RTX 4090 24 GB on Amazon

ASIN B0BG94PS2F. Amazon Associates link. The live price is on that page, not locked in here.

Related guides

Sources

A correction or a hardware question goes to the contact page.