Which Local Model Can Actually Drive Claude Code? The `ollama launch` VRAM Map

Short answer: ollama launch claude is one command, and on a 24GB card with muse-glimmer:30b (an 18GB download, 128K context, Apache 2.0) it is a genuinely usable setup. On 12GB you drop to a 9B-class driver like qwen3.5:9b (6.6GB). On 16GB, gemma4:12b (7.6GB) leaves room to breathe. Below 8GB, don't. And the number that decides all of this is not parameters — it's that Ollama defaults to 4k context under 24GiB of VRAM while every agent harness in its own docs asks for 64k minimum.

That last sentence is the whole page. Everything else is detail.

---

What fits, by VRAM tier

Sizes below are the exact download sizes published on each model's Ollama tag page, checked 18 August 2026. They are the weights only. Add KV cache for whatever context you set — see the next section, because on the small tiers it is the difference between working and thrashing.

The "Context on card" and "Our read" columns are our judgement, not vendor figures — the download sizes and maximum context windows are. Treat the first two columns as fact and the last two as a starting point to test against.

A few notes that matter more than the table:

• The -mlx tag is bigger, not smaller. 21GB against 18GB for the standard tag. People assume the Apple build is leaner; it is not. Ollama added DFlash and image input support to the MLX path in 0.32.7.
• Muse Glimmer has a lot of tags. 30b-nvfp4 is 17GB, 30b-q4_K_M is 18GB, 30b-mxfp8 is 33GB, 30b-bf16 is 57GB, and the DFlash variants add a few GB each. The plain muse-glimmer tag resolves to the 18GB build — that is the one you want unless you know why you want otherwise.
• Qwen3.5 goes both smaller and much larger. 0.8B (1.0GB) through 122B-A10B (81GB), all at 256K context. If you're mapping a whole fleet, our Ollama RAM/VRAM table and the 24GB VRAM coding picks cover the wider lineup.

---

Context is the real bottleneck, not parameters

Ollama's documentation states its default context lengths outright: under 24GiB VRAM you get 4k, 24-48GiB gets 32k, and 48GiB or more gets 256k. The same page then says tasks that need large context — "web search, agents, and coding tools" — should be set to at least 64000 tokens.

Read those two facts together and the failure mode is obvious. A 12GB card boots your agent with a 4k window. Claude Code loads a system prompt, a directory tree and one file, and it's already out of room. Nothing errors. The agent just gets stupid.

Raise it explicitly:

OLLAMA_CONTEXT_LENGTH=64000 ollama serve

Then confirm nothing spilled to CPU:

ollama ps

Check the PROCESSOR column reads 100% GPU and the CONTEXT column reads what you asked for. If it says something like "62%/38% CPU", your tokens-per-second just fell off a cliff and the agent will feel broken even though it technically works. Drop to a smaller tag or a lower context before you conclude the model is bad — our Ollama RAM/VRAM table has the smaller options.

Ollama's docs repeat the 64k floor in three separate integration pages we read — Codex ("requires a larger context window", at least 64k), OpenClaw ("at least 64k tokens if using local models"), and Claude Code ("for larger repositories, set the context length to 64k or higher"). When three vendor docs independently name the same number, treat it as the requirement, not a suggestion.

---

The commands

Every command below is copied from Ollama's own docs or release notes as of 18 August 2026.

Install a driver model and start Claude Code against it:

ollama pull muse-glimmer
ollama launch claude --model muse-glimmer

The same shape works across harnesses:

ollama launch pi --model muse-glimmer        # minimal, extensible coding agent
ollama launch openclaw --model muse-glimmer  # personal assistant, messaging bridges
ollama launch hermes --model muse-glimmer    # self-improving skills + memory
ollama launch codex                          # OpenAI Codex CLI
ollama launch dsh --model qwen3.8            # DeepSeek Harness (developer preview)

Useful flags, all documented:

ollama launch                     # interactive picker — the current, authoritative list
ollama launch claude --config     # configure without starting
ollama launch codex --restore     # remove Ollama's launch profile and generated catalog
ollama launch dsh -- --port 3081  # pass arguments through to the harness

And the non-interactive form for scripts, Docker or CI, straight from the Claude Code integration doc:

ollama launch claude --model gemma4:cloud --yes -- -p "how does this repository work?"

--yes skips the selectors, pulls the model if needed, and requires --model.

If you'd rather not use the launcher at all, the manual wiring for Claude Code is three variables:

export ANTHROPIC_AUTH_TOKEN=ollama
export ANTHROPIC_API_KEY=""
export ANTHROPIC_BASE_URL=http://localhost:11434
claude --model qwen3.5

That manual path is what our older Claude Code offline with Ollama guide describes, and it still works — the launcher just does it for you and adds a model picker.

---

Which harness for which job

ollama launch is not one integration; the Ollama repo's docs directory currently carries integration pages for 26 tools. The five that matter for local coding, with what the vendor docs actually say about each:

Pi's tool set is the honest tell about the whole category. Four tools and an extension system means the harness is thin and the model is doing the work. Which is exactly why the model choice above matters more than the harness choice.

Per-tool walkthroughs live in their own guides: Codex CLI, OpenCode, Crush, Kilo Code, Zed, Goose and Hermes. This page is the model-selection layer above them.

---

What we did and did not verify

Being blunt about this, because a wrong VRAM number costs you an evening:

Verified against primary sources, 18 August 2026:

• Release tags and dates from the GitHub release records: v0.32.0 (11 July 2026), v0.32.7 (10 August 2026), v0.32.8 (10 August 2026), v0.32.11 (14 August 2026), v0.32.14 (15 August 2026).
• Every model size and context figure from the model's own tag listing on ollama.com.
• Every command from Ollama's docs directory or a release body.
• Muse Glimmer's licence (Apache 2.0), 128K context, 30B parameters and image input from its model card.
• The default-context-by-VRAM figures from Ollama's context-length documentation.

Not verified, and therefore not claimed anywhere on this page: tokens-per-second on any card, task completion rates, and any head-to-head "which model finishes the refactor" comparison. We have not run a controlled benchmark across these harnesses, and publishing invented throughput numbers is how buying guides get people to waste money. When we have measured numbers, they'll appear here with the methodology attached.

Vendor claims, labelled as such: Ollama states its MLX engine gives "state-of-the-art performance on Apple Silicon" for Muse Glimmer, and describes Muse Glimmer as built for always-on local agents. Those are Ollama's and Meta's words, not a measurement of ours.

---

Honest limitations

Where this setup will disappoint you:

1. "Offline" is conditional. First launch installs the harness over npm or a vendor script. Web search inside Claude Code, Pi, OpenClaw and DeepSeek Harness routes through Ollama's hosted search and needs ollama signin. The inference is local; the convenience layer is not.
2. Context eats your card. Covered above, and it is the single most common reason a local agent underperforms the same model in a chat window.
3. DeepSeek Harness is a preview. Ollama's own doc says the upstream configuration may change between releases. Don't build a workflow on it yet.
4. The launcher writes config. ollama launch codex uses a dedicated profile and regenerates a model catalog; ollama launch dsh writes ~/.ollama/launch/dsh/settings.yaml. Ollama's docs say it leaves your own ~/.dsh/settings.yaml, profiles, sessions and credentials alone — but know that files are being written on your behalf.
5. A 30B local model is not a frontier cloud model. It is roughly free, private, and yours. Set expectations accordingly and give it smaller, better-scoped tasks than you'd hand a hosted agent.
6. Ollama is shipping fast. Nine tagged releases between 4 and 15 August 2026 alone (v0.32.6 through v0.32.14). Anything in this space, including this page, has a short half-life — check ollama launch with no arguments for the live list.

---

Verdict

If you have 24GB of VRAM or a 48GB Mac, local agentic coding stopped being a science project this summer. One ollama launch claude and an 18GB model download gets you there, and the setup friction that used to fill a whole blog post is now handled by the launcher.

If you have 12-16GB, you can absolutely run this — just go in knowing you're driving a 9B-to-12B model with a hand-raised 64k window, and that the ceiling is real. If you have 8GB, spend the evening on smaller-model workflows instead and save agentic coding for your next card.

And whatever tier you're on: set the context length before you judge the model. Most "local agents are useless" takes are 4k-context takes.

---

FAQ