Best Open Source Self-Hosted LLMs for Coding in 2026

The gap between proprietary and open source AI models for coding is narrowing fast. A year ago, self-hosting an LLM for development meant settling for significantly worse performance than cloud-based alternatives like GPT-5.4 or Claude. In 2026, the best open source models are closing in on proprietary leaders across independent benchmarks like Artificial Analysis and LiveBench, and some even outperform them on specific tasks like code generation and completion.

Whether you’re a solo developer who wants to keep code off third-party servers, a startup looking to cut API costs, or an enterprise with strict data compliance requirements, self-hosted open source LLMs have become a genuinely viable option for professional software development. In this guide, we’ll cover the best open source models you can self-host for coding, the tools to deploy them, and the hardware you need to get started.

Open Source vs Proprietary: How Close Is the Gap?

Before diving into individual models, it’s worth understanding where open source stands. We use Artificial Analysis as the primary lens here because it’s independent and, unlike LiveBench, already covers the June 2026 releases (MiniMax M3, Kimi K2.7 Code, DeepSeek-V4-Pro-Max). Its Intelligence Index aggregates provider-reported and benchmark-derived signals into one number. We cross-check it against SWE-Bench Pro (coding-specific) and LiveBench (contamination-aware, but still missing the newest models) further down. The snapshot below is July 2026.

Artificial Analysis Intelligence Index (July 2026, open weights)

The Self-host column is the approximate VRAM just for 4-bit (INT4/Q4) weights and the smallest 80GB-class multi-GPU box that fits them, from each model’s community and vendor deployment notes ( GLM-5.2, MiniMax M3, DeepSeek V4, Kimi K2.7). Two caveats: add headroom on top for the KV cache, which balloons at these models’ 256K-1M context lengths, and none of these run on a single consumer GPU. If that’s your budget, skip to the smaller models below - Devstral Small 2 (24B) and Qwen 3.6 27B fit a single RTX 4090 or a 32GB Mac. You can also trade GPUs for large CPU RAM with GGUF builds (a 2-bit Kimi quant, for instance, runs on a 384GB DDR5 workstation), at much lower speed.

On Artificial Analysis, GLM-5.2 is the clear open-weight leader at 51.1, ahead of the tight 42-44 cluster of MiniMax M3, DeepSeek V4 Pro, Kimi K2.6/K2.7, and MiMo-V2.5-Pro. The gap to the frontier is real but narrow: the best proprietary coding model, Claude Opus 4.8, scores 56 (the amber bar in the banner) - about five points ahead of GLM-5.2, not the chasm it was a year ago. On AA’s real-world agentic benchmark (GDPval-AA v2), GLM-5.2 scores 1524, effectively level with GPT-5.5 xHigh (1514). At the small end, Qwen3.6-35B-A3B trails at 32 but is the one model here you can actually run on a laptop. The two cross-checks below tell the same story.

LiveBench (secondary cross-check, July 2026)

LiveBench is contamination-aware, which makes it a useful sanity check, but it hasn’t yet scored the three June releases, so they’re absent here. Among the models it does cover, GLM-5.2 again leads on both metrics.

LiveBench Agentic Coding Average

On LiveBench too, GLM-5.2 leads the open-source set: 73.33 Agentic Coding Avg (above) and 79.65 Coding Avg - the agentic score also beats the proprietary GPT-5.4 Thinking xHigh Effort (70.00). Kimi K2.6 Thinking is second on both (58.33 agentic, 78.57 coding).

For the latest scores and full model list, visit the LiveBench leaderboard directly.

SWE-Bench Pro (coding-specific cross-check)

SWE-Bench Pro is the coding-specific benchmark that does cover the June releases, so it’s the best head-to-head on code tasks alone. GLM-5.2 tops the open-weight field here too, with the June releases clustered just behind.

DeepSeek-V4-Pro-Max sits lower on SWE-Bench Pro but leads open weights on the older SWE-Bench Verified at 80.6% (tied with Gemini 3.1 Pro) and posts 93.5% on LiveCodeBench, so its ranking depends heavily on which benchmark you weight. Several of these are self-reported by the vendor; treat them as directional until LiveBench and other independent evaluations catch up.

Best Open Source LLMs for Coding

New in June 2026: MiniMax M3, Kimi K2.7 Code, and DeepSeek-V4-Pro-Max

Three open-weight releases landed in June 2026, after the last version of this guide. None are on LiveBench yet, so their numbers are vendor and third-party reported for now.

MiniMax M3 (MiniMax)

MiniMax M3 shipped on June 1, 2026 as an open-weight model combining frontier coding performance, a 1M-token context, and native multimodal input in a single architecture (weights followed within about ten days of the API launch). Its headline feature is MSA (MiniMax Sparse Attention), which partitions the KV cache into blocks so each block is read only once, which MiniMax says delivers more than 4x faster attention than Flash-Sparse-Attention style implementations and much faster prefill at long context. It’s reported at roughly 428B parameters, served through vLLM and SGLang. Vendor benchmarks report 59.0% SWE-Bench Pro, 66.0% Terminal-Bench 2.1, and 74.2% MCP-Atlas.

Kimi K2.7 Code (Moonshot AI)

Kimi K2.7 Code is Moonshot’s coding-specialized follow-up to K2.6, released on June 12, 2026 under a Modified MIT license. It keeps the ~1T total / 32B active MoE design and 256K context, adds a MoonViT vision encoder, and uses roughly 30% fewer reasoning tokens than K2.6, which directly lowers the cost of long agent runs. Moonshot reports 58.6 SWE-Bench Pro and a hallucination-rate drop from about 65% to 39%, plus a +21.8% gain on its own Kimi Code Bench v2 over K2.6. Every published K2.7 number is first-party so far; there are no independent leaderboard results yet.

DeepSeek-V4-Pro-Max (DeepSeek)

DeepSeek shipped DeepSeek-V4 on April 24, 2026 in two variants: V4-Pro (1.6T total / 49B active) and V4-Flash (284B total / 13B active), both 1M context, MIT-licensed. The higher-effort V4-Pro-Max configuration leads open weights on the older SWE-Bench Verified at 80.6% (tied with Gemini 3.1 Pro) and posts 93.5% on LiveCodeBench, though on the stricter SWE-Bench Pro it lands at 55.4, behind GLM-5.2. On LiveBench, V4-Pro scores 69.99 Coding and 56.67 Agentic Coding.

The April 2026 wave that these build on is also still worth a look: GLM-5.1 (section #1), Kimi K2.6 (section #2), Qwen3.6 (35B-A3B and 27B, Apache 2.0), and Xiaomi’s MiMo-V2.5-Pro (section #5).

1. GLM-5.2 / GLM-5.1 (Z.AI) - Leading Open-Source Coding Family

The GLM-5 family from Z.AI has become the most competitive open-source option for long-horizon coding tasks. The newest release, GLM-5.2 (June 2026), extends its predecessor’s 200K context to a full 1M tokens and is the first open-weight model to beat GPT-5.5 on SWE-Bench Pro. The prior GLM-5.1 (200K context, 58.4 SWE-Bench Pro, up to 8-hour long-horizon execution) remains a solid LiveBench-listed fallback for teams that have already benchmarked it.

What makes this family particularly noteworthy is its training infrastructure. The GLM-5 generation was trained on 100,000 Huawei Ascend 910B chips rather than NVIDIA GPUs - a significant milestone for non-NVIDIA AI hardware. Z.AI also introduced a novel reinforcement learning infrastructure called “Slime” that reduced hallucination rates from 90% to 34%, and GLM-5.2 adds anti-hack mechanisms in RL training specifically for coding agents.

GLM-5.2’s architecture introduces IndexShare, which reuses the sparse attention indexer across every four sparse attention layers - cutting per-token FLOPs by 2.9x at 1M context length without sacrificing quality. An improved MTP layer increases speculative decoding acceptance length by up to 20%. It also adds two selectable thinking modes: Max for maximum reasoning depth and High for a better latency/quality tradeoff.

On coding benchmarks, GLM-5.2 scores 79.65 Coding Avg and 73.33 Agentic Coding Avg on LiveBench - the highest open-source results in this guide on both metrics, and the agentic coding score beats every proprietary model in the table. On SWE-Bench Pro it posts 62.1 (above GPT-5.5’s 58.6 and GLM-5.1’s 58.4), 81.0 on Terminal-Bench 2.1 (vs GLM-5.1’s 63.5), 74.4 on FrontierSWE (vs Claude Opus 4.8’s 75.1 and GPT-5.5’s 72.6), and 76.8 on MCP-Atlas (vs GPT-5.5’s 75.3). Artificial Analysis independently ranks it the top open-weights model too: 51 on its Intelligence Index (ahead of MiniMax M3 and DeepSeek V4 Pro at 44, and Kimi K2.6 at 43) and 1524 on the real-world GDPval-AA v2 agentic benchmark, effectively level with GPT-5.5 xHigh.

Key Specs - GLM-5.2 (June 2026)

• Architecture: MoE, 753B total / 40B active parameters
• Context Window: 1M tokens
• License: MIT
• SWE-Bench Pro: 62.1 (self-reported by Z.AI; beats GPT-5.5 at 58.6)
• Terminal-Bench 2.1: 81.0 (self-reported by Z.AI)
• FrontierSWE: 74.4 (self-reported by Z.AI)
• MCP-Atlas: 76.8 (self-reported by Z.AI)
• LiveBench Coding Avg: 79.65
• LiveBench Agentic Coding Avg: 73.33 (highest open-source in this guide; beats GPT-5.4 Thinking xHigh at 70.00)
• Self-hosting: vLLM (v0.23.0+), SGLang (v0.5.13.post1+), KTransformers, Transformers; weights on Hugging Face and ModelScope; multi-GPU required

2. Kimi K2.6 / K2.7 Code (Moonshot AI) - Second on LiveBench Coding, Best Agentic Stability

Kimi K2.6 is Moonshot AI’s flagship open-weight coding model line, listed on Moonshot’s latest research timeline on April 20, 2026. On the May 12, 2026 LiveBench snapshot, Kimi K2.6 Thinking is the second-strongest open-source model in this guide with 78.57 Coding Avg and 58.33 Agentic Coding Avg, behind only GLM-5.2. It’s the pick when you value agentic stability - recoverable failure modes and consistent tool calling across long sessions.

Moonshot’s technical write-up for K2.6 describes a larger agent-swarm setup, with up to 300 sub-agents and around 4,000 coordinated reasoning/execution steps for complex workflows. In practice, this is aimed at repo-level tasks where planning, tool use, and verification have to run over long trajectories.

On the official model card, Moonshot reports K2.6 scores of 58.6 (SWE-Bench Pro), 80.2 (SWE-Bench Verified), 66.7 (Terminal-Bench 2.0), and 89.6 (LiveCodeBench v6). These are vendor-reported numbers, but they position K2.6 as one of the strongest open-weight coding options currently available.

Moonshot has since shipped Kimi K2.7 Code (June 12, 2026), a coding-specialized variant on the same ~1T / 32B active MoE base. It matches K2.6’s SWE-Bench Pro (58.6, vendor-reported) while using about 30% fewer reasoning tokens and cutting the hallucination rate from roughly 65% to 39%. If you’re standing up Kimi for agentic coding today, K2.7 Code is the version to pull; K2.6 Thinking is still the one with the independent LiveBench numbers quoted above.

Key Specs

• Architecture: MoE, ~1T total / 32B active parameters
• Context Window: 256K tokens
• License: Modified MIT (commercial use allowed)
• SWE-Bench Pro: 58.6 (self-reported by Moonshot AI)
• SWE-Bench Verified: 80.2 (self-reported by Moonshot AI)
• Terminal-Bench 2.0: 66.7 (self-reported by Moonshot AI)
• LiveCodeBench v6: 89.6 (self-reported by Moonshot AI)
• LiveBench Coding Avg (May 12, 2026): 78.57
• LiveBench Agentic Coding Avg (May 12, 2026): 58.33
• Self-hosting: Recommended via vLLM or SGLang; production requires 2x H100 80GB or 4x A100 80GB, 512GB RAM

3. DeepSeek V3.2 - Coding 75.69, Agentic 46.67

DeepSeek V4 (covered above) is the current flagship, but V3.2 is still worth listing as the best cost-to-quality baseline. It scores 75.69 on LiveBench Coding Average and 46.67 on Agentic Coding, with 671 billion total parameters (37 billion active), a Mixture of Experts architecture, a 160K context window, and 73.1% on SWE-bench Verified. It’s MIT-licensed.

The model’s API pricing is remarkably low at roughly $0.27 to $0.55 per million tokens, making it one of the most cost-effective options even before considering self-hosting. For local deployment, the smaller DeepSeek Coder models (6.7B) run comfortably on consumer hardware through Ollama or LM Studio, while the full V3.2 requires enterprise-grade infrastructure.

Key Specs

• Architecture: MoE, 671B total / 37B active parameters
• Context Window: 160K tokens
• License: MIT
• LiveBench Coding Avg: 75.69
• LiveBench Agentic Coding Avg: 46.67
• SWE-bench Verified: 73.1% (self-reported by DeepSeek)
• Self-hosting: Full model requires multi-GPU setup; smaller Coder variants run on consumer GPUs via Ollama

4. Devstral 2 (Mistral AI) - Coding 66.79, Agentic 43.33

Devstral 2 from Mistral AI is a 123 billion parameter model specifically designed for agentic software engineering. It scores 66.79 on LiveBench Coding Average and 43.33 on Agentic Coding. Released in December 2025, it scores 72.2% on SWE-bench Verified with a 256K context window, making it one of the most capable code-focused models available. Mistral describes it as 7x more cost-efficient than Claude Sonnet and 5x smaller than DeepSeek V3.2 while remaining competitive in benchmarks.

What makes the Devstral family compelling for self-hosting is the smaller sibling, Devstral Small 2 (24B parameters), which scores an impressive 68% on SWE-bench Verified. That’s remarkable for a model that runs on a single RTX 4090 or a Mac with 32GB of RAM. It also supports image inputs and comes with Apache 2.0 licensing, making it one of the most permissive options available.

Mistral also offers Vibe CLI, an open source terminal coding assistant powered by Devstral, giving you a ready-made development workflow out of the box.

Key Specs (Devstral 2)

• Parameters: 123B
• Context Window: 256K tokens
• License: Modified MIT
• LiveBench Coding Avg: 66.79
• LiveBench Agentic Coding Avg: 43.33
• SWE-bench Verified: 72.2% (self-reported by Mistral AI)
• Self-hosting: Multi-GPU recommended for full model

Key Specs (Devstral Small 2)

• Parameters: 24B
• Context Window: 128K tokens
• License: Apache 2.0
• SWE-bench Verified: 68.0% (self-reported by Mistral AI)
• Self-hosting: Single RTX 4090 or Mac with 32GB RAM

5. MiMo-V2.5-Pro (Xiaomi) - 78.9% SWE-Bench, 68.4% TerminalBench

MiMo-V2.5-Pro is Xiaomi’s latest open-weight model, released on April 22, 2026. It’s a 1.02T total parameter MoE model with 42B active parameters and a 1M token context window - broadly comparable in scale to DeepSeek-V4 Pro. Weights are on Hugging Face and ModelScope under the MIT license.

MiMo-V2.5-Pro is not currently listed on the LiveBench leaderboard, so a direct side-by-side comparison with the other models in this guide isn’t possible yet. On vendor-reported benchmarks, it posts 78.9% on SWE-Bench Verified and 68.4% on TerminalBench 2.0. The TerminalBench score is the highest in this guide, slightly ahead of Kimi K2.6’s 66.7. These are self-reported numbers from Xiaomi; treat them as directional until LiveBench or independent evaluations confirm them.

The architecture uses a hybrid attention design that interleaves local sliding window attention with global attention at a 6:1 ratio, which Xiaomi says cuts KV-cache memory usage by roughly 7x compared to full attention at long contexts. Three lightweight Multi-Token Prediction modules enable a 3x inference speedup. For self-hosting, SGLang is the recommended inference engine; the model requires a significant multi-GPU setup similar to other ~1T MoE models in this guide.

Xiaomi describes MiMo-V2.5-Pro as a major step forward from MiMo-V2-Pro for agentic and software engineering tasks, with support for workflows involving more than 1,000 sequential tool calls.

Key Specs

• Architecture: MoE, 1.02T total / 42B active parameters
• Context Window: 1M tokens
• License: MIT
• SWE-Bench Verified: 78.9% ( self-reported by Xiaomi)
• TerminalBench 2.0: 68.4% ( self-reported by Xiaomi)
• LiveBench Coding Avg: Not listed on LiveBench
• LiveBench Agentic Coding Avg: Not listed on LiveBench
• Self-hosting: SGLang or vLLM; multi-GPU setup required (similar footprint to DeepSeek V4 Pro)

6. Qwen3-Coder (Alibaba) - Best Agentic CLI Tool

The Qwen3-Coder family from Alibaba represents one of the most comprehensive open source coding model lineups available. The flagship model features 480 billion parameters with a Mixture of Experts design, and Alibaba describes it as “our most agentic code model to date.” There’s also a smaller 30B variant (3B active) for resource-constrained environments.

The more recent Qwen3-Coder-Next (80B total, 3B active) pushes the envelope further with hybrid attention combined with MoE, trained with large-scale reinforcement learning specifically for agentic tasks. It scores 70.6% on SWE-bench Verified, an impressive result for a model with only 3B active parameters.

Alibaba also provides Qwen Code, an open source terminal coding agent optimized for Qwen3-Coder models. This gives developers a Claude Code or Aider-like experience powered entirely by open source infrastructure.

The broader Qwen ecosystem also includes Qwen 2.5 Coder (available in sizes from 0.5B to 32B), which remains one of the best mid-range options. The 32B Instruct variant scores 73.7 on the Aider benchmark (comparable to GPT-4o) and is readily available through Ollama.

Key Specs

• Architecture: MoE, up to 480B total parameters
• License: Apache 2.0
• SWE-bench Verified: 70.6% (Qwen3-Coder-Next, self-reported by Alibaba)
• Self-hosting: Qwen 2.5 Coder 32B runs on consumer hardware via Ollama; larger variants require multi-GPU

7. Llama 4 (Meta) - Largest Context Window (10M)

Llama 4 from Meta continues to be the most widely deployed open source model family, with over 650 million total downloads and roughly 9% of enterprise production workloads running on Llama variants. The Llama 4 family released in April 2025 includes Scout (109B total, 17B active, 10M context window), Maverick (400B total, 17B active, 1M context), and the announced but unreleased Behemoth (~2T total, 288B active).

While Llama 4 isn’t specifically a coding model, its massive context windows and multimodal capabilities (text and image input across 12 languages) make it highly versatile for development workflows. The code-specific Llama 4 Coder variant brings improved code generation, debugging, and completion accuracy.

The main caveat is licensing: Llama’s license does not meet the OSI Open Source Definition and includes restrictions for companies with very large user bases. For most developers and smaller organizations, this is a non-issue, but it’s worth noting compared to the MIT or Apache 2.0 licenses of other models on this list.

Key Specs

• Architecture: MoE, up to 400B total / 17B active (Maverick)
• Context Window: Up to 10M tokens (Scout)
• License: Llama Community License (restrictions for very large companies)
• Self-hosting: Scout and Maverick available via Ollama, vLLM; smaller variants run on consumer hardware

8. StarCoder 2 (BigCode / Hugging Face) - Most Auditable Training Data

StarCoder 2 is a collaboration between Hugging Face and ServiceNow under the BigCode project. Available in 3B, 7B, and 15B sizes, it was trained on 3.3 to 4.3 trillion tokens from The Stack v2, covering 619 programming languages. It uses Grouped Query Attention with a 16K context window.

StarCoder 2’s standout quality is its data transparency. Every training data source is documented with Software Heritage Identifiers (SWHIDs), making it the most auditable coding model available. This matters for enterprises concerned about IP and licensing compliance. The 15B model matches or outperforms CodeLlama 34B (a model twice its size), demonstrating strong efficiency.

While it doesn’t compete with the larger MoE models on raw benchmarks, StarCoder 2 remains an excellent choice for teams that need a lightweight, well-documented coding model they can run on modest hardware.

Key Specs

• Sizes: 3B, 7B, 15B
• Context Window: 16K tokens
• License: OpenRAIL (fully transparent training data)
• Self-hosting: Runs on consumer hardware via Ollama; 3B variant works on laptops

Honorable Mentions

Several other open source models deserve recognition for specific strengths:

• IBM Granite Code - Available from 350M to 34B parameters under Apache 2.0, trained on 116 programming languages with license-permissible data. Granite 4.0 introduces hybrid Mamba-2/transformer architecture using 70% less memory. Best choice for enterprise compliance.

• NVIDIA Nemotron-Cascade 2 - A 30B MoE with only 3B active parameters that achieves Gold Medal-level performance on competitive programming benchmarks (IMO, IOI, ICPC) with 20x fewer parameters than comparable models. Remarkable efficiency.

• Yi-Coder - From 01.AI, available in 1.5B and 9B sizes with 128K context and Apache 2.0 license. Yi-Coder 9B scores 85.4% on HumanEval, on par with DeepSeek Coder 33B at a fraction of the size.

• Qwen 3.5 - Released February 2026 with a 397B MoE model, featuring unified vision-language capabilities and support for 201 languages. One of the top-ranked open-weight models across multiple benchmarks.

How to Use These Models with a Coding Agent

If you want a Claude Code or Aider-style workflow with self-hosted models, one of the easiest setups is OpenCode + Ollama. This combination gives you a local coding agent with a simple terminal workflow and no cloud dependency.

Easiest Setup: OpenCode + Ollama

If you’re using Ollama’s built-in Applications flow, the setup is even simpler. The current Qwen 3.6 Ollama page lists a direct OpenCode launch command.

Step 1: Install Ollama

curl -fsSL https://ollama.com/install.sh | sh

Step 2: Install OpenCode

curl -fsSL https://opencode.ai/install | bash

Step 3: Launch OpenCode directly through Ollama Applications

ollama launch opencode --model qwen3.6:35b-a3b

Step 4: Open your project and start working

Once OpenCode starts, point it at your repository and use it like any other terminal coding agent for explaining code, refactoring files, writing tests, or implementing features.

If you want a smaller local footprint, Ollama also provides smaller Qwen 3.6 variants (for example 27B-class options). Check the live Ollama model page for currently available tags.

Why This Setup Works Well

• Fastest setup path because Ollama can launch OpenCode directly as an application
• Runs fully local with no separate model gateway to configure
• Easy to scale up or down by swapping the Ollama model tag based on your hardware

How to Self-Host These Models Locally

Once you’ve picked a model, you need the right tools and hardware to run it. We’ve covered this extensively in our previous guides:

• How to Self-Host Any LLM - Step by Step Guide - A complete walkthrough covering installation, model download, quantization, GPU setup, and connecting to your development tools.
• Top 5 Local LLM Tools and Models - A detailed comparison of Ollama, vLLM, llama.cpp, LM Studio, and other self-hosting tools with hardware requirements and performance benchmarks.

Quick Decision Guide

Conclusion

The open source LLM landscape for coding has matured dramatically, and GLM-5.2 still holds the bar it raised in June. It leads the open-weight field on every source we tracked: Artificial Analysis Intelligence Index (51.1), SWE-Bench Pro (62.1), and LiveBench (79.65 Coding, 73.33 Agentic - the agentic number beats every proprietary model in that table). The June 2026 wave - MiniMax M3, Kimi K2.7 Code, and DeepSeek-V4-Pro-Max - clusters just behind, though several of their numbers are still vendor-reported, so their standing is directional until the independent leaderboards finish scoring them.

For most developers, the practical recommendation is to start with Qwen 3.6 27B or Devstral Small 2 on local hardware, then move to GLM-5.2, Kimi K2.7 Code, MiniMax M3, or DeepSeek-V4-Pro-Max if you need top-tier agentic performance and have enterprise GPUs. DeepSeek V3.2 remains a strong cost-to-quality baseline, and MiniMax M3 is worth a look when you specifically need a 1M-token context with native multimodal input.

The 44% of organizations that cite data privacy as their top concern with LLM adoption now have no reason to hold back. Self-hosted open source models are production-ready for coding, and the gap with proprietary alternatives continues to shrink with each new release.