Open coding models you can host yourself, and the GPU each needs

The best local LLM for coding is the one that fits the GPU you can rent, so I sort the open coding models by the memory they need and quote what each vendor publishes on its model card. One 48 GB card holds Devstral Small 2 at 25.79 GB. One 80 GB card holds Qwen3.8-27B or gpt-oss-120b. A 2x H200 VM holds Mistral Medium 3.5 at 133.61 GB, and a full 8-GPU VM holds GLM-5.3, DeepSeek V4.1-Flash and Kimi K2.7 Code. I do not rank them, because every card reports its own benchmark with its own harness.

How I grouped the open source coding models

I group by the GPU a model needs because that is the first limit a rented VM puts on you, and inside a group I list the models by size and not by score. People who search for the best open source coding model or the best coding LLM usually want one name, and no card here supports one. Sizes are the summed safetensors files, counted once where the Mistral and gpt-oss repositories hold two copies. The smallest-VM column is calculated as 92% of the VM's memory minus the checkpoint, so it leaves out activations and CUDA graphs. The score column names the benchmark as the card names it, and every score is vendor-reported.

The harnesses differ, so a number in one row says nothing about a number in another. The GLM and MiniMax cards run DeepSWE with mini-SWE-agent, the DeepSeek V4-Flash-0731 card uses DeepSeek's own harness, the Qwen3.8-27B card names Terminus for Terminal Bench 2.1, and the MiniMax-M3 card names Claude Code as the scaffolding for SWE-Bench Verified. Every size, score and license on this page was checked on 12 November 2026.

Best local LLM for coding on one 48 GB card

Three models are small enough for the 48 GB tier as released. Devstral Small 2 is a 24B dense model from Mistral in FP8. MiMo-V2.6-Distill-Qwen-9B is Qwen3.5-9B fine-tuned on MiMo-generated data, which Xiaomi publishes as a starting point for research. gpt-oss-20b is OpenAI's 21B mixture-of-experts model in MXFP4. The FP8 builds of Qwen3.8-27B at 30.87 GB and Qwen3-Coder-30B-A3B at 31.18 GB also fit a 48 GB Ada card, and the next group lists them. FP8 runs natively on the Ada cards in this tier, as the Qwen3-Coder page explains.

Coding models for one 80 to 141 GB card

An 80 GB card holds the 27B and 30B models in BF16 and gpt-oss-120b in MXFP4, and the Qwen page and the gpt-oss page carry their commands. The Qwen3-Coder page covers the 30B model, whose card shows its scores only as an image and points to Qwen's blog. An H200 at 141 GB also holds the FP8 build of Qwen3-Coder-Next at 80.38 GB with about 49 GB left, and that model's scores are in the next group.

Coding models for 2 to 4 GPUs

Checkpoints of 120 to 360 GB need two to four GPUs, and most rows below fit 2x or 4x H200. The Mistral page has the Medium 3.5 command and the MiMo page the MiMo-V2.6-Flash command, while the DeepSeek page, the GLM page and the MiniMax-M3 page cover V4-Flash, GLM-5.3-Flash and MiniMax-M3.1-Flash. Mistral's offline models page also recommends Mistral Small 4 for local use, and its card has no coding score in the text, only charts. Pick a 2x or 4x H200 offer whose Interconnect column says NVLink, and see the multi-GPU page for how the GPUs share a model.

The Modified MIT License of Devstral 2 and Mistral Medium 3.5 grants no rights to a company whose monthly revenue exceeds 20 million dollars, and the Qwen Community License asks a business that runs a Model as a Service or an AI Work Assistant, which it defines as an independent product made mainly for AI-assisted coding or office work, to get a separate license from Qwen for commercial use, with an exception for internal use. Both matter if you build a coding product, so read the files.

Coding models that need a full 8-GPU VM

These checkpoints run from 510 GB to 1,561 GB, and the frontier models page compares them on one table of GPU memory. The Kimi K3 page covers the largest two and the B300 stock problem, and the frontier models page has the Kimi K2.7 Code command. The GLM, DeepSeek, MiniMax-M3, MiMo and Mistral pages cover their own rows. Anything past 8 GPUs goes through the cluster quote form.

What a local coding agent needs from the server

A local coding agent is a terminal or editor tool that calls a model server you run, and the server has to do two things. It needs tool calling, because the agent reads files and runs commands through tool calls. With vLLM that means --enable-auto-tool-choice and a --tool-call-parser for the model. vLLM's tool calling page lists qwen3_xml for Qwen3-Coder, mistral for the Mistral models, openai for gpt-oss and mimo for MiMo-V2.6, and the model pages above give the flag for each family. It also needs a long context, because the agent keeps files and tool output in it.

The Claude Code guide and the OpenCode guide cover the clients, the server flags and the context settings. DeployGPU's private app links do not accept API keys or Bearer tokens, so a coding tool reaches the server through an SSH tunnel to a plain GPU VM, as the vLLM guide shows.

Memory for a long context

The weights are a fixed cost and the KV cache grows with every token of context, so a card with little left over holds a short session. Calculated from each model's config.json, BF16 cache costs 160 KiB per token for Devstral Small 2, 96 KiB for Qwen3-Coder-30B-A3B, 352 KiB for Devstral 2 and Mistral Medium 3.5, and 22.5 KiB for Mistral Small 4 and MiMo-V2.6-Flash. A 128,000-token session is then about 19.5 GiB for Devstral Small 2, 11.7 GiB for Qwen3-Coder-30B-A3B, 43 GiB for Medium 3.5 and 2.7 GiB for Small 4. The 18 GB that a 48 GB card has left after Devstral Small 2 holds about 112,000 tokens, and an FP8 KV cache halves each figure where the model supports it. The KV cache explainer shows the method, and the engine's startup log has the real number.

Kimi Code and Qwen Code are agents, not models

Kimi Code is Moonshot AI's coding agent. Its page describes an AI coding agent powered by Kimi K3.1 Pro with a CLI, a VS Code extension and a desktop app, sold under paid plans. The CLI is open source under the MIT license, and its provider documentation lists type openai for the OpenAI Chat Completions protocol and compatible services with a base_url you can override, so it can reach a vLLM server through a tunnel. The Kimi weights you can host are in the table above, and Kimi K2.7 Code is the one its card calls coding-focused.

Qwen Code is the Qwen team's open-source coding agent for the terminal, licensed Apache-2.0. Its README says it supports "Any third-party provider or local model (Ollama / vLLM)". Mistral Vibe works the same way, and Mistral's docs say the Vibe Code CLI "supports any model served behind an OpenAI-compatible API".