The best Ollama model for coding depends on the work you need it to perform, the memory available to it and the tools around it. This guide identifies model options and deployment checks. It is a documentation-based comparison, not a set of locally measured rankings.
Correction — October 7, 2026: The previous version described RTX 4090, RTX 3090 and M3 Max tests, 50+ runs and per-language pass rates without supporting first-party run records in the reviewed material. Those claims and the resulting rankings have been removed. HumanEval, MMLU, LiveCodeBench and SWE-bench results are no longer mixed into one score table.
What Ollama Provides
Ollama can run downloaded models locally and expose them through its API. It also supports cloud features, so selecting Ollama does not by itself establish that an entire workflow stays on your computer. Check the selected model, remote services and IDE configuration. The Ollama FAQ distinguishes local and cloud operation.
The API is not Docker-compatible. Ollama has its own API plus compatibility with parts of the OpenAI API. Client support, tool calling and completion behavior need to be checked for the actual model and integration.
Running your own model can avoid a hosted per-token bill for that inference. It still incurs hardware, electricity and maintenance costs. Cloud tools or remote calls used by an agent can introduce separate charges.
Coding Model Options to Evaluate
Each linked catalog entry below identifies a model or tag. The order is not a performance ranking, and no row is a guarantee that the model fits a particular GPU.
| Model option | Catalog tag | Useful evaluation focus |
|---|---|---|
| Qwen3.6 27B | qwen3.6:27b | Repository edits and tool-assisted coding |
| Devstral 24B | devstral:24b | Agentic workflows with multiple files |
| Qwen2.5-Coder 32B | qwen2.5-coder:32b | Code generation, completion and repair |
| Qwen3-Coder-Next | qwen3-coder-next | Coding-agent workloads with a larger memory budget |
| DeepSeek-Coder-V2 Lite | deepseek-coder-v2:16b | Coding tasks using the smaller Coder-V2 variant |
| DeepSeek-R1 32B distill | deepseek-r1:32b | Reasoning-heavy tasks, assessed against executable tests |
| Phi-4 | phi4:14b | Smaller general-model baseline for logic-heavy tasks |
| Llama 3.3 70B | llama3.3:70b | General assistance plus coding on larger-memory systems |
These candidates come from the original article’s scope; the list is not an exhaustive inventory of current releases. Check tags and model IDs before testing, particularly when a family contains several generations or a latest alias can change.
Qwen3.6 and Devstral
Ollama’s Qwen3.6 catalog lists the 27B variant with a 256K advertised context window, not the 128K limit previously stated here. The Devstral catalog lists devstral:24b; the earlier devstral:24b-small command was not the verified tag. See their respective catalog entries above.
For both candidates, test whether your agent correctly handles tool responses, edits all required files and runs the project’s checks. An agent’s result depends on the orchestration and tools as well as the model. This article has no controlled evidence supporting the former tool-call accuracy or relative throughput percentages.
Qwen2.5-Coder and DeepSeek Variants
Keep the code-specialized Qwen2.5-Coder family distinct from general Qwen2.5 checkpoints. Likewise, distinguish DeepSeek-Coder-V2 Lite from its larger counterpart, and an R1 distill from the full R1 model. A benchmark for one checkpoint cannot be assigned to another just because the names share a family prefix.
Choose tasks that reflect the intended use: isolated function generation, repairing an existing module, or completing an issue through an agent. A correct function does not establish that the same setup can navigate a repository, install dependencies and produce a working patch.
Qwen3-Coder-Next Is Not a 16 GB All-GPU Model
Qwen3-Coder-Next has 80B total parameters and 3B active per token. The checked Ollama catalog lists its Q4_K_M artifact at approximately 52 GB. The earlier recommendation for fully resident 4-bit use on a 16 GB GPU was incorrect. Source: Ollama catalog.
Offloading some weights to system memory is a different deployment. It requires enough total memory and has its own performance characteristics. Active parameters reduce per-token computation; they do not remove the other experts from storage.
Phi-4 and Llama 3.3
These provide general-model alternatives to evaluate against code-focused candidates. Llama 3.3 is a 70B model; the checked Ollama catalog lists its default artifact at about 43 GB. Leave additional room for context and runtime allocations. A general knowledge score such as MMLU is not a HumanEval coding score, so the earlier substitution has been removed.
Hardware: Inspect the Artifact Before Choosing a GPU
Weight size, runtime memory and GPU memory are different quantities. A model file that is smaller than your GPU’s nominal capacity may still fail to fit once context, cache and runtime overhead are included. Conversely, a model may run partly on the CPU without being a viable interactive coding assistant for your latency target.
Check the exact quantization and desired context. Leave memory for the operating system and other applications on a unified-memory machine. Do not interpret 24 GB on a laptop or GPU as 24 GB exclusively available to the model.
Use ollama show to inspect the model and ollama ps after loading it to check actual CPU/GPU placement. These commands are documented by the Ollama project. Fixed tokens-per-second estimates need a measured hardware and workload configuration; this guide does not supply such measurements.
A Minimal Setup Example
Install Ollama using the official download instructions. The following is an illustrative local setup, not a benchmark run:
ollama pull qwen2.5-coder:7b
ollama show qwen2.5-coder:7b
ollama run qwen2.5-coder:7b
ollama ps
The smaller tag is an example, not a universal recommendation. Confirm the available variant and its requirements in the Qwen2.5-Coder catalog.
curl http://localhost:11434/api/chat -d '{
"model": "qwen2.5-coder:7b",
"messages": [{"role": "user", "content": "Write a Python binary search function and tests for empty input and missing values."}],
"stream": false
}'
Review the generated code before running it. If the server is not running, start it using the method appropriate to your installation. Avoid launching another server against an already occupied port.
For IDE use, follow that IDE or extension’s current Ollama integration documentation. A native provider may expect http://localhost:11434, whereas an OpenAI-compatible integration may expect a /v1 base URL. Do not assume an old configuration file format or that every IDE can reach a loopback endpoint.
How to Produce a Defensible Local Comparison
Before selecting a winner, choose representative tasks and define pass/fail rules. Include the existing code, dependencies and tests the model is expected to work with. Keep task inputs and tool permissions consistent across candidates.
Record:
- Exact model ID, quantization, runtime version and context configuration.
- Hardware, loaded memory placement and concurrent workloads.
- Task definition, prompts, tool harness and retry limits.
- Functional success, test results and human corrections needed.
- First-token delay, total completion time and peak memory.
- Failed and timed-out runs, not only successful examples.
Report public benchmark results separately from your own runs. Do not compare numbers from different metrics as if they share a scale. A leaderboard result without the same harness, precision and inference budget does not validate a local deployment.
FAQ
Which is the best Ollama model for coding?
This article does not establish a tested winner. Shortlist candidates that fit your resources and compare them on the coding workflow you actually need.
Can a useful coding model run on 8 GB of VRAM?
Smaller or more heavily quantized checkpoints may fit, depending on context and runtime overhead. Check the artifact and measure the actual allocation; there is no blanket fit guarantee for an entire parameter class.
Does MoE mean a model requires less memory than a dense model?
Not necessarily. Compare total weight storage, cache and placement. A small active-parameter count is not a memory specification.
Is a local model free to operate?
There may be no hosted token charge for local inference, but hardware, power, time and any remote tools still have costs.