A coding model that lives on your card can keep proprietary code off someone else’s server and remove the per-query bill, but only if the model is bilingual in “fill in this line” and “fix this issue across four files.” Most chat models do one or the other; the open-weight coding models do both, and the cluster has changed substantially since the 2026-03-17 lineup. Every figure below was pulled from the Hugging Face model card, the model’s own GitHub repo, or the live Ollama tag page on 2026-09-22.
Quick reference
| Job | Model | Ollama tag (size) | Licence | SWE-bench Verified |
|---|---|---|---|---|
| Smallest FIM-capable coder | Qwen2.5-Coder-0.5B | qwen2.5-coder:0.5b (398 MB) | Apache 2.0 | not on card |
| 1.5B for low-RAM laptops | Qwen2.5-Coder-1.5B | qwen2.5-coder:1.5b (986 MB) | Apache 2.0 | not on card |
| Phone-class chat-completion | Qwen2.5-Coder-3B | qwen2.5-coder:3b (1.9 GB) | Apache 2.0 | not on card |
| 8 GB sweet spot, FIM | Qwen2.5-Coder-7B | qwen2.5-coder:7b (4.7 GB) | Apache 2.0 | not on card |
| 16 GB tier, agentic | Qwen2.5-Coder-14B | qwen2.5-coder:14b (9.0 GB) | Apache 2.0 | not on card |
| 16 GB tier, agentic | Devstral Small 2 | devstral:24b (14 GB) | Apache 2.0 | 68.0 |
| 24 GB tier, agentic | KAT-Coder-V2.5-Dev | (no Ollama tag) | Apache 2.0 | 69.40 |
| 24 GB tier, large dense | Qwen2.5-Coder-32B | qwen2.5-coder:32b (20 GB) | Apache 2.0 | not on card |
| 24 GB tier, MoE flagship | Qwen3-Coder-30B-A3B | qwen3-coder:30b (19 GB) | Apache 2.0 | not on card |
| 24 GB tier, MoE agentic | GLM-4.7-Flash | (no Ollama tag) | MIT | 59.2 |
| 64 GB+ tier | Qwen3-Coder-Next | qwen3-coder-next (52 GB) | Apache 2.0 | 70.6 |
| Frontier MoE | Qwen3-Coder-480B-A35B | qwen3-coder:480b (290 GB) | Apache 2.0 | not on card |
| Frontier MoE, retired | GLM-4.7 | (Ollama tag retired) | MIT | 73.8 |
| Frontier MoE, retired | Kimi K2-Instruct-0905 | (Ollama tag retired) | Modified MIT | 69.2 |
| Non-production FIM | Codestral-22B | (no Ollama tag) | MNPL-0.1 | not on card |
Two notes before the details. First, SWE-bench Verified scores are not on every card: the Qwen2.5-Coder series and the 30B-A3B / 480B Qwen3-Coder variants list Agentic Coding, Browser-Use, and Tool-Use numbers in the Qwen blog but not SWE-bench Verified on the model pages I read. Second, Ollama retired both glm-4.7 (July 15 2026) and kimi-k2 (June 16 2026) - the tags still resolve on the library page but show “no models have been pushed,” so they are listed for completeness but cannot be pulled today.
Two jobs, two architectures
Autocomplete (fill-in-the-middle, FIM). The model receives text before and after the cursor and has to produce the missing line. The classic test is HumanEval; most strong coding models now score in the 85-90% range, which makes HumanEval a poor differentiator. FIM support itself is the harder filter: a chat model can answer a coding question without ever having been trained on FIM, so it cannot do autocomplete.
Agentic coding. The model receives a real GitHub issue, navigates a repository, edits multiple files, runs tests, and revises. The test is SWE-bench Verified (and the harder SWE-bench Pro and SWE-bench Multilingual). This is where the MoE coding models - KAT-Coder, Qwen3-Coder, GLM-4.7-Flash, GLM-4.7 - separate from the dense Qwen2.5-Coder line, because the model has to keep state across many tool calls.
Some models do both. Qwen2.5-Coder ships FIM and instruction tuning at every size from 0.5B to 32B and remains the most consistent autocomplete-capable family for low-VRAM cards. Qwen3-Coder and KAT-Coder dropped FIM support in favour of agentic strength; pairing one of them with a small Qwen2.5-Coder model on the same card, one for autocomplete and one for the heavier agent turns, is a common pattern.
Pick by tier
Under 4 GB. qwen2.5-coder:0.5b (398 MB) and qwen2.5-coder:1.5b (986 MB) are FIM-capable at sizes that fit anything, but expect limited agentic ability. qwen2.5-coder:3b (1.9 GB) is the more useful pick at this tier if you have the RAM; the Qwen2.5-Coder family is the only coding line that goes this small without losing FIM support entirely. The no-GPU cluster entry point covers what to expect on CPU-only hardware.
4 to 6 GB. qwen2.5-coder:7b (4.7 GB) is the standout at this budget, Apache 2.0 and FIM-capable. This is also where the 8 GB VRAM tier guide and the quantisation quality page apply, because the same card will also be running a chat or RAG model.
9 to 14 GB. Two choices diverge here. qwen2.5-coder:14b (9.0 GB) is the dense agentic coder; it has FIM and is Apache 2.0. devstral:24b (14 GB on the Ollama library page; the HF model card lists the Small 2 24B Instruct 2512 at the same parameter class) is the agentic specialist from Mistral: SWE-Bench Verified 68.0, Apache 2.0, 256K context, but Mistral shipped no FIM for the Small 2 release. The Ollama devstral and devstral-small-2 tags are both still listed (1.2M and 1.0M pulls respectively at time of writing); check the tag landing page before recommending it for new deployments so you can verify the specific tag you mean. The 16 GB tier guide covers the broader picks for this card size.
19 to 20 GB. The dense tier. qwen2.5-coder:32b (20 GB) is the dense Apache 2.0 128K-context option for cards that do not want an MoE. qwen3-coder:30b (19 GB) is the MoE flagship in this band: 30.5B total / 3.3B active, Apache 2.0, 262K native context extendable to 1M via YaRN, non-thinking mode only. The Qwen3-Coder family supports agentic tool-use, and the Qwen blog documents three integration paths (Qwen Code CLI, Claude Code proxy, Cline with the qwen3-coder-plus model name). GLM-4.7-Flash (30B/3B MoE, MIT, SWE-bench Verified 59.2) is the closest competitor at the same MoE footprint, but Ollama has not pushed a tag and the vLLM/SGLang serving path requires nightly wheels. The 24 GB tier guide covers the chat and RAG alternatives.
Above 24 GB. Three frontiers exist. qwen3-coder-next (52 GB Q4_K_M, Apache 2.0) is the 80B/3B MoE follow-up with SWE-bench Verified 70.6, Terminal-Bench 2.0 36.2, SWE-bench Pro 44.3, and a 262K native context - the strongest single-card coding model in the open-weight family that I could verify against a primary source. qwen3-coder:480b (290 GB Q4_K_M, Apache 2.0) is the 480B/35B flagship with 256K context extendable to 1M via YaRN - the same architectural ceiling as the 30B-A3B but at a different deployment scale, and the closest local Qwen3-Coder build to a hosted Sonnet-4 class. GLM-4.7 (MIT, 358B, SWE-bench Verified 73.8) and Kimi K2-Instruct-0905 (Modified MIT, 1T/32B MoE, SWE-bench Verified 69.2) score higher but both have been retired from Ollama (glm-4.7 retired 2026-07-15, kimi-k2 retired 2026-06-16) and require multi-GPU FP8 serving - they remain strong picks for llama.cpp or vLLM clusters but are not single-card consumer hardware. The 32 GB tier guide covers the hardware ceiling.
Licences, where the traps are
The coding-model family is mostly permissive. Qwen2.5-Coder (all six sizes), Qwen3-Coder-30B-A3B, Qwen3-Coder-Next, Qwen3-Coder-480B-A35B, Devstral Small 2, and KAT-Coder-V2.5-Dev are all Apache 2.0. GLM-4.7 and GLM-4.7-Flash are MIT. Kimi K2-Instruct-0905 is the Modified MIT variant Moonshot uses for its larger models, with usage-count thresholds above the stock MIT terms - read the LICENSE file before deploying commercially.
The one outright non-permissive entry is Codestral-22B-v0.1, Mistral’s earlier coding specialist. It is the only model on this page that ships a custom Mistral MNPL-0.1 (Mistral AI Non-Production License), which prohibits production deployment; the HF card says “does not have any moderation mechanisms” and asks the community to add guardrails. For any commercial use, replace it with devstral:24b (Apache 2.0, SWE-bench Verified 68.0) at the same parameter class. The open-model licences guide covers the broader patterns.
Autocomplete versus agentic in practice
A coding model in front of a real repository is most useful if it can both fill in the next line and execute a multi-step issue. The two jobs want different things from the same weights:
- FIM wants a fast, dense model that loads fast and stays on the card. Qwen2.5-Coder at 7B or 14B is the strongest match on consumer hardware.
- Agentic work wants a MoE with a long context and tool-call training. Qwen3-Coder-30B-A3B at 19 GB (262K native context, 1M via YaRN, Apache 2.0) is the smallest mainstream pick that holds up; Qwen3-Coder-Next at 52 GB (SWE-bench Verified 70.6) is the strongest single-card frontier.
The common production pattern is to load both: a small Qwen2.5-Coder for the editor’s inline suggestion and a larger MoE coder for the agent turn. Ollama keeps both loaded if VRAM allows; the Ollama runners comparison covers which runners expose num_parallel and keep_alive knobs.
How to actually run one
Two flags matter most. For Qwen3-Coder, the model card recommends temperature=0.7, top_p=0.8, top_k=20, repetition_penalty=1.05, and max_new_tokens=65,536; it is non-thinking only. For Qwen2.5-Coder, the original model card recommends lower temperature (0.1 to 0.3) for FIM. The GGUF, AWQ, GPTQ and MLX comparison covers which format to pull, and what quantisation costs you covers the quality loss, which on a coding model is more visible than on chat because errors compound across file edits.
For tool-use setups, the Qwen3-Coder integration guide documents three paths: the Qwen Code CLI (forked from Gemini Code), Claude Code via a proxy API or claude-code-router, and Cline with the OpenAI-compatible provider pointed at a local Ollama endpoint using qwen3-coder-plus as the model name. Each path assumes the upstream Ollama or vLLM server is already running; the household-serving guide and the team-serving guide cover the network plumbing.
The bottom line
For most consumer cards, the answer is qwen2.5-coder:14b (9.0 GB, Apache 2.0, FIM-capable) for autocomplete paired with qwen3-coder:30b (19 GB, Apache 2.0, 262K context) for agentic turns - both fit on a 24 GB card with headroom for context. On a 16 GB card, qwen2.5-coder:7b plus devstral:24b (14 GB, SWE-bench Verified 68.0, Apache 2.0) is the strongest verifiable pairing; both Ollama tags are currently listed as live at time of writing, but tag state changes quickly so confirm on the Ollama library page before deployment. Above 24 GB, qwen3-coder-next (52 GB Q4_K_M, SWE-bench Verified 70.6, Apache 2.0) is the strongest single-card open-weight coding model I could verify against a primary source. The only non-permissive licence in the family is Codestral-22B’s MNPL-0.1; everything else on the table is Apache 2.0, MIT, or Modified MIT and can be deployed commercially once the Modified-MIT usage threshold is read against the LICENSE file.