A current ranking of open-source and open-weight coding models, reviewed by exact artifact, benchmark evidence, license, weight access, and practical deployment.
Quick answer
GLM 5.2 is our current overall pick for long-horizon coding agents, repository work, and terminal tasks. Choose Qwen3 Coder 480B-A35B Instruct when Apache-2.0 terms and repository-scale context are the priority.
Read the selection methodologyCurrent ranking
This is an editorial ranking, not a disguised family-recency list or a composite of incompatible vendor benchmarks. Every entry is tied to an exact reviewed artifact.
Z.ai
Best overall for long-horizon coding agents
Our current top editorial pick because its official release combines strong coding-agent results, a 1M-token context window, published weights, and permissive MIT terms.
Qwen
Best Apache-2.0 model for repository-scale coding
A coding-specific checkpoint with published weights, Apache-2.0 terms, native 262K context, and first-party agentic coding guidance.
Moonshot AI
Best for long-horizon agentic coding under custom terms
A long-horizon agentic-coding model with published weights, Modified-MIT license terms, a native 1M-token context window, and vendor-reported terminal-agent coding results.
Mistral AI
Best reviewed option for a high-memory local workstation
The official release documents a 24B Apache-2.0 checkpoint that can run on a single RTX 4090 or a Mac with 32 GB memory.
Poolside
Efficient MoE pick for long-horizon agentic coding
A compact MoE coding model with published weights, the permissive OpenMDW-1.1 license, a native 1M-token context window, and vendor-reported results competitive with larger models.
Methodology
Models qualify only when we can review an exact coding-relevant artifact, its weight status, and its license. The order is editorial because coding benchmarks still use different agents, prompts, budgets, contexts, and attempt counts.
Benchmark and artifact snapshot reviewed . The hosted catalog refreshes independently and does not change the editorial review date.
Published or gated weights, reviewable terms, and clear coding relevance.
Agent and coding results, weighted by source quality and harness transparency.
License terms, weight access, hosted availability, and practical self-hosting.
Workflow fit and evidence quality decide rank; incompatible scores are not averaged.
Evidence snapshot
Results below retain their original harness labels. They are evidence for each checkpoint, not a normalized cross-vendor league table.
| Model | License | Weight access | Published evidence | Harness note |
|---|---|---|---|---|
| GLM 5.2 | MIT | available | SWE-bench Pro: 62.1% | OpenHands with the release prompt and a 400K context window |
| Qwen3 Coder 480B-A35B Instruct | Apache-2.0 | available | SWE-bench Pro: 38.7% | Scale AI public evaluation listed on the model card |
| Kimi K3 | Modified MIT (Kimi K3 License) | available | Terminal-Bench 2.1: 88.3% | Kimi Code harness, max reasoning effort |
| Devstral Small 2505 | Apache-2.0 | available | SWE-bench Verified: 46.8% | OpenHands release evaluation |
| Laguna S 2.1 | OpenMDW-1.1 | available | SWE-Bench Multilingual: 78.5% | Poolside release evaluation |
Vendor-reported results can vary with scaffold, prompt, tool permissions, context, compute budget, and number of attempts. Open the linked source before making a procurement or deployment decision.
Most current coding leaders are more precisely described as open-weight models. Weight access alone does not settle rights to modify, redistribute, fine-tune, or use an artifact commercially.
Compare definitions and license checksRun a reviewed artifact locally when your hardware fits, use a hosted Kilo route, or connect a provider key. Large MoE checkpoints usually need server-class inference even when their active parameter count looks small.
This page owns evidence-reviewed open-model recommendations. Use these focused pages for live, free, new-release, or hardware-specific questions.
Compare all models by Kilo usage, KiloBench, price, speed, and context.
See zero-price hosted models with comparable KiloBench evidence.
Track recent releases chronologically without treating recency as rank.
Choose exact artifacts, runtimes, quantizations, and memory tiers.
Review the exact artifact and terms before deploying. A lab name or similar model slug is never treated as evidence that another checkpoint is open.
Released 2026-06-17 · reviewed 2026-07-22
Released 2025-07-22 · reviewed 2026-07-22
Released 2026-07-27 · reviewed 2026-07-30
Released 2025-05-21 · reviewed 2026-07-22
Released 2026-07-21 · reviewed 2026-07-30
GLM 5.2 is our current overall editorial pick for long-horizon coding agents, repository work, and terminal tasks. The best choice still depends on your license, hardware, latency, and workflow requirements.
Not always. Open-weight means the model weights are available under stated terms. Open-source AI has broader requirements around the freedoms and information needed to use, study, modify, and share the system. This ranking uses artifact-level classifications and shows the license for every model.
Devstral Small 2505 is the strongest reviewed local-workstation option in this ranking. Local fit depends on the exact artifact, quantization, runtime, context length, and available memory.
This is an editorial ranking based on coding-agent evidence, benchmark quality, license and deployment freedom, practical availability, efficiency, and maintenance. Vendor results are not merged into a single composite score when harnesses differ.
The current Kilo model feed matched these reviewed artifacts: GLM 5.2, Qwen3 Coder 480B-A35B Instruct, Kimi K3, Laguna S 2.1. Availability can change, so unmatched entries link to their primary model source instead.
Live zero-price hosted models and tested free winners.
Hardware-first local model recommendations.
Chronological verified release tracker.
Definitions, licensing, and procurement checklist.
Production data from 10,643 AI code reviews across 13 models.
Try current hosted models, connect your own provider keys, or route local inference through Kilo alongside 500+ model options.