WhichQwenCodermodelshouldyourunonOllama?

Which Qwen Coder models exist today?

Qwen has shipped three coder generations, and all three are still downloadable. The table below is built from the Hugging Face model cards and the Ollama library pages as fetched on October 26, 2026.

Three details trip people up:

• “Qwen coder next” means Qwen3-Coder-Next. It was released in early February 2026, per the Hugging Face collection dates, and its model card describes it as built for coding agents and local development.
• There is no “Qwen4 Coder” in the lineup. The Qwen Hugging Face organization lists no such model as of October 26, 2026, so treat third-party listings of one with suspicion.
• Qwen’s newer general models are not Coder-branded. Qwen3.8-27B, a dense 27B Apache 2.0 model, claims strong coding results on its own model card, but those are the publisher’s self-reported numbers. This guide covers the Coder line only. For how it stacks up against other coding models, see the best local LLMs for coding.

What does MoE vs dense mean for running these locally?

A mixture-of-experts (MoE) model stores all its weights in memory but only computes a few of them per token, so it needs the memory of a big model and the speed of a small one. Qwen3-Coder-30B-A3B holds 30.5B parameters but activates 3.3B per token, according to its model card. Qwen2.5-Coder is dense: every parameter is used on every token.

In practice that gives you a trade:

• Memory is set by total parameters. The 30B MoE costs roughly what a 30B dense model costs to load, and the 19 GB Ollama download is the floor, not the total.
• Speed is set by active parameters. Fewer active parameters means less computation per token than a dense 32B, though real throughput depends on your hardware and runtime.
• Quality isn’t predicted by either number. Don’t pick on parameter count alone. Qwen’s own benchmark tables are self-reported, and this site hasn’t run its own tests on these models yet.

Which size fits your hardware?

One developer who pulled the 51 GB Qwen3-Coder-Next through Ollama onto an RTX 5090 with 32 GB of VRAM measured about 12 tok/s, because roughly half the layers landed in system RAM (HN comment, Feb 2026). That is the failure this section is about. Pick the largest model whose download size fits in your VRAM or unified memory with several GB to spare for the context cache. The file size is a hard lower bound on memory use, and a long context adds to it. Sizes below are Ollama download sizes from the library pages.

LM Studio lists a minimum of 15 GB of system memory for its Qwen3-Coder-30B build, which is lower than Ollama’s 19 GB download, so the two defaults are evidently different builds. See local LLM hardware requirements for how to size a GPU or Mac for any of these.

Quantization matters as much as the model name. Ollama publishes 199 Qwen2.5-Coder tags across q2_K up to fp16. The official Qwen3-Coder-Next GGUF repository lists Q4_K_M at 48.4 GB, Q8_0 at 84.8 GB and F16 at 159 GB. Lower bit-widths shrink the file and cost some accuracy, and how much depends on the model, so verify on your own code.

How much context do you actually get?

Qwen3 Coder models support 262,144 tokens natively, but Ollama will not give you that by default. Ollama chooses its default context from your VRAM: 4k tokens under 24 GiB, 32k from 24 to 48 GiB, and 256k at 48 GiB or more, per its documentation. The same page recommends at least 64,000 tokens for coding tools and agents.

Set it explicitly, and confirm with ollama ps:

OLLAMA_CONTEXT_LENGTH=64000 ollama serve

Other facts from the model cards:

• Qwen3-Coder-30B-A3B can be extended to 1M tokens with YaRN, and Qwen3-Coder-Next is validated up to 131,072 with YaRN scaling in its GGUF card.
• Qwen2.5-Coder’s default config is capped at 32,768 tokens. YaRN reaches 131,072, but the card warns static YaRN in vLLM can hurt short-text performance.
• Every extra token of context costs memory on top of the weights. If a model runs out of memory, the Next card suggests cutting context to 32,768 first.

What do developers report in practice?

Community reports agree on the shape: the 30B MoE is the practical sweet spot, the 80B Next is a memory-and-offload problem, and agent use depends on the client. These are individual reports on different hardware, not our tests.

Three practical lessons come out of them:

• Fit the whole model in fast memory. The 5090 report is the clearest case: an 80B model that spills to system RAM runs at CPU speed no matter how fast the GPU is.
• Tool calling needs the right parser. Users of the 30B model in Roo Code, Cline and opencode reported inconsistent agent behavior until the server used the right tool-call parser, for example --enable-auto-tool-choice --tool-call-parser qwen3_coder on vLLM (Level1Techs forum, Aug 2025). Unsloth’s Qwen3-Coder-Next guide also says llama.cpp fixes for looping and tool-call parsing landed after launch, so update llama.cpp and re-download older GGUFs (Unsloth guide).
• Expect a ceiling on agent work. One developer who keeps coming back to Qwen3-Coder 30B on a 20 GB workstation GPU describes it as decent in isolation but weaker on multi-step, context-heavy agentic work than hosted frontier models (HN comment, Oct 2025). For the wider trade-off, see local LLM vs Claude for coding.

Base vs instruct: which do you need?

Use an instruct model for chat, agents and editor assistants, and a base model only for raw completion or your own fine-tuning. The Qwen2.5-Coder base card says “we do not recommend using base language models for conversations,” and points to post-training or fill-in-the-middle (FIM) tasks instead.

The Qwen3 generation changes the picture:

• Qwen3-Coder 30B and 480B are instruct only. The Qwen3-Coder README labels them as instruct models for chatting, and says FIM is supported in every version, so the instruct builds already handle completion.
• Qwen3-Coder-Next has both. Qwen3-Coder-Next-Base exists on Hugging Face, but it has had far fewer downloads than the instruct model (about 1.6 thousand versus about 479 thousand at the time of writing), and the Ollama library has no base tag.
• Qwen2.5-Coder has both at every size. Ollama carries -base and -instruct tags for all six sizes. For tab-completion in an editor, the small base builds are the natural fit.
• None of the Qwen3-Coder models think out loud. The cards say they only support non-thinking mode and do not emit <think> blocks. Don’t build prompts that expect a reasoning trace.

Recommended sampling differs by generation: temperature 0.7, top_p 0.8, top_k 20 for Qwen3-Coder-30B-A3B, and temperature 1.0, top_p 0.95, top_k 40 for Qwen3-Coder-Next. Check the card for your exact model.

How do you run Qwen Coder in Ollama, LM Studio and llama.cpp?

Ollama is the shortest path: one pull, one run. All three runtimes load the same GGUF weights, so the choice is about interface, not quality.

# Ollama: the default 30B MoE (19 GB)
ollama pull qwen3-coder:30b
ollama run qwen3-coder:30b

# Ollama: Qwen3-Coder-Next, 4-bit (52 GB)
ollama run qwen3-coder-next:q4_K_M

# Ollama: the dense 14B for a 16 GB GPU
ollama run qwen2.5-coder:14b

# Straight from Hugging Face's official GGUF repo
ollama run hf.co/Qwen/Qwen3-Coder-Next-GGUF:Q4_K_M

For llama.cpp, the official GGUF card gives a serve command in the form ./llama-server -hf Qwen/Qwen3-Coder-Next-GGUF:Q4_K_M, plus a llama-cli line that uses --jinja, -fa on and --no-context-shift. Copy it from the card rather than from a blog, since flags change between releases. In LM Studio, search for the model name, or use the Qwen3-Coder-30B page’s “Use Model in LM Studio” link, which also offers MLX builds at 4 to 8 bits for Apple silicon.

What are the license terms?

Almost all Qwen Coder models are Apache 2.0, which allows commercial use, modification and redistribution. The Hugging Face license field reads apache-2.0 for the Qwen3-Coder models, Qwen3-Coder-Next and its Base and GGUF repos, and Qwen2.5-Coder at 0.5B, 1.5B, 7B, 14B and 32B.

The exception is Qwen2.5-Coder-3B, whose card lists qwen-research rather than Apache 2.0. Read that license before shipping it in a product.

Apache 2.0 requires you to pass on the license text, mark files you changed, and keep notices if you redistribute. It also ends your patent license if you sue over the Work. It does not add an acceptable-use policy, which is a difference from Llama’s license. Hosting providers may add their own terms.

What should you know about abliterated Qwen Coder variants?

Abliterated Qwen Coder models are community re-uploads with the refusal behavior edited out, not Qwen releases. A Hugging Face search on October 26, 2026 found several for Qwen3-Coder, including huihui-ai’s Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated and an abliterated Qwen3-Coder-Next, plus many GGUF conversions by third parties.

Three things to keep in mind:

• They are unofficial. The huihui-ai card calls its method “a crude, proof-of-concept implementation” and says the model “has not undergone rigorous safety optimization.” Quality loss is possible and unmeasured here.
• The upstream license still applies. The cards list Apache 2.0, but you remain responsible for what you do with the output, and for using it on code and systems you are authorized to work on.
• Check whether you need one. Test the official weights on your own workload first. Read what abliterated models are and how to run them in Ollama before swapping out the official weights.

For head-to-head questions about DeepSeek Coder, see DeepSeek Coder vs Qwen Coder.

FAQ