{"slug":"best-open-weight-llm","title":"Best open-weight LLM","question":"What are the best open-weight LLMs for building applications in 2026?","category":"Models","url":"https://modelsagree.com/best/best-open-weight-llm","updated":"2026-07-15","models":["ChatGPT","Claude","Gemini","Grok"],"consensus":"2 of 4 models rank GLM-5.2 the top pick","disagreement":"Claude picks DeepSeek-V3.2; Gemini picks DeepSeek-V4-Pro","combined":[{"rank":1,"product":"GLM-5.2","domain":"z.ai","score":12,"appearances":3,"modelRanks":{"ChatGPT":1,"Gemini":4,"Grok":1},"reason":"Best overall balance of frontier-grade coding, reasoning, tool use, long-horizon agent work, and low-cost hosted access; narrowly leads DeepSeek-V4-Pro for application development."},{"rank":2,"product":"DeepSeek-V4-Pro","domain":"deepseek.com","score":9,"appearances":2,"modelRanks":{"ChatGPT":2,"Gemini":1},"reason":"Offers frontier-level intelligence for coding, reasoning, and agentic workflows with a dual Thinking/Non-Thinking mode under a permissive MIT license, while dramatically reducing KV cache usage."},{"rank":3,"product":"DeepSeek-V3.2","domain":null,"score":5,"appearances":1,"modelRanks":{"Claude":1},"reason":"Frontier-adjacent quality under a true MIT license, with the best cost-to-capability ratio in the open ecosystem — strong reasoning (R1 lineage distilled in), solid coding and tool use, and it is servable via every major inference provider or self-hostable, so practitioners get GPT-4-class output without vendor lock-in; ranked first assuming the builder can use hosted inference rather than running the full MoE themselves"},{"rank":4,"product":"Kimi K2.5","domain":null,"score":4,"appearances":1,"modelRanks":{"Grok":2},"reason":"Native multimodal agentic capabilities with strong vision+coding, agent swarms, and frontier-competitive benchmarks (high SWE-Bench Verified, GPQA); efficient MoE (1T total/32B active); open weights under Modified MIT for flexible app integration."},{"rank":5,"product":"Llama 4 Maverick","domain":null,"score":4,"appearances":1,"modelRanks":{"Gemini":2},"reason":"Meta's flagship 400B MoE model offers excellent generalist reasoning and coding capabilities with a 1M token context, highly optimized for fast inference with only 17B active parameters."},{"rank":6,"product":"Qwen3","domain":null,"score":4,"appearances":1,"modelRanks":{"Claude":2},"reason":"The most complete open family for application builders — Apache 2.0, sizes from sub-1B to 235B MoE sharing one behavior profile, first-rate multilingual coverage, strong function calling, and the deepest fine-tuning/quantization ecosystem, so one stack scales from edge to server; near-tie with DeepSeek, losing only on peak reasoning quality"},{"rank":7,"product":"gpt-oss-120b","domain":"openai.com","score":3,"appearances":2,"modelRanks":{"ChatGPT":5,"Claude":4},"reason":"Apache 2.0 with the best capability-per-GPU in the open field — MXFP4 quantization lets it run on a single 80GB GPU with strong reasoning and adjustable effort levels, making it the most realistic self-host option for teams that must keep data on their own hardware"},{"rank":8,"product":"DeepSeek V4","domain":null,"score":3,"appearances":1,"modelRanks":{"Grok":3},"reason":"Exceptional efficiency (MoE with low active params) and 1M+ context at competitive cost/performance; strong on math/reasoning/agentic coding; MIT license and proven self-hosting value for scalable app backends."},{"rank":9,"product":"Kimi K2","domain":null,"score":3,"appearances":1,"modelRanks":{"Claude":3},"reason":"The strongest open-weight agentic model — its 1T-parameter MoE (32B active) leads open models on tool-calling, multi-step agent tasks, and coding benchmarks, making it the best drop-in for agent frameworks where reliability of function calls is the bottleneck; modified-MIT license is permissive in practice"},{"rank":10,"product":"Kimi K2.7 Code","domain":null,"score":3,"appearances":1,"modelRanks":{"ChatGPT":3},"reason":"Particularly strong for repository-scale coding and autonomous software-engineering workflows, making it an excellent foundation for coding agents and application builders."},{"rank":11,"product":"Qwen3-Coder-480B-A35B","domain":null,"score":3,"appearances":1,"modelRanks":{"Gemini":3},"reason":"Extremely strong specialized performance in multi-turn software engineering tasks and coding agents, backed by a fully permissive Apache 2.0 license for unrestricted commercial use."},{"rank":12,"product":"Qwen3 / Qwen3.6","domain":null,"score":2,"appearances":1,"modelRanks":{"Grok":4},"reason":"Versatile family with strong coding/reasoning variants, hybrid modes, Apache 2.0 license, and excellent multilingual/long-context support; practical for diverse app development with good efficiency across sizes."},{"rank":13,"product":"Qwen3.6-35B-A3B","domain":null,"score":2,"appearances":1,"modelRanks":{"ChatGPT":4},"reason":"The practical-value leader: strong multilingual reasoning, coding, and tool use with only about 3B active parameters, Apache 2.0 licensing, and realistic local deployment."},{"rank":14,"product":"Gemma 4","domain":null,"score":1,"appearances":1,"modelRanks":{"Gemini":5},"reason":"Google's lightweight unified multimodal open-weights model supporting text, image, video, and audio on Apache 2.0, running efficiently on consumer-grade workstation GPUs."},{"rank":15,"product":"GLM-4.6","domain":null,"score":1,"appearances":1,"modelRanks":{"Claude":5},"reason":"The best open model specifically for coding-agent applications — near-Claude-Sonnet coding performance, strong long-context tool use in harnesses like Claude Code and Cline, MIT license, and cheap hosted inference; ranked on the assumption the practitioner is building developer-facing or code-generating products"},{"rank":16,"product":"Llama 4","domain":null,"score":1,"appearances":1,"modelRanks":{"Grok":5},"reason":"Robust ecosystem, tooling, and community support for integration; native multimodal MoE with solid general performance and commercial-friendly license (up to usage limits); reliable for production apps."}],"perModel":{"ChatGPT":[{"rank":1,"product":"GLM-5.2","reason":"Best overall balance of frontier-grade coding, reasoning, tool use, long-horizon agent work, and low-cost hosted access; narrowly leads DeepSeek-V4-Pro for application development.","fix":"Its roughly 753B-weight footprint makes practical self-hosting a multi-GPU or datacenter undertaking."},{"rank":2,"product":"DeepSeek-V4-Pro","reason":"Near-tied with GLM-5.2, with exceptional reasoning, coding, long-context processing, agentic execution, a permissive license, and unusually strong API value.","fix":"The enormous 1.6T-parameter model is unrealistic for typical practitioners to self-host."},{"rank":3,"product":"Kimi K2.7 Code","reason":"Particularly strong for repository-scale coding and autonomous software-engineering workflows, making it an excellent foundation for coding agents and application builders.","fix":"Its coding specialization makes it less compelling than the top generalists for broad knowledge, writing, or mixed enterprise workloads."},{"rank":4,"product":"Qwen3.6-35B-A3B","reason":"The practical-value leader: strong multilingual reasoning, coding, and tool use with only about 3B active parameters, Apache 2.0 licensing, and realistic local deployment.","fix":"It cannot match the frontier models above on the hardest long-horizon reasoning and agentic tasks."},{"rank":5,"product":"gpt-oss-120b","reason":"Mature open tooling, Apache 2.0 licensing, strong reasoning and function calling, and single-80GB-GPU operation make it a dependable customizable production choice.","fix":"Its text-only capabilities and older performance ceiling now trail newer open-weight leaders."}],"Claude":[{"rank":1,"product":"DeepSeek-V3.2","reason":"Frontier-adjacent quality under a true MIT license, with the best cost-to-capability ratio in the open ecosystem — strong reasoning (R1 lineage distilled in), solid coding and tool use, and it is servable via every major inference provider or self-hostable, so practitioners get GPT-4-class output without vendor lock-in; ranked first assuming the builder can use hosted inference rather than running the full MoE themselves","fix":"The full model is a massive MoE that is impractical to self-host on modest hardware, and there is no strong small-size family — teams needing on-device or single-GPU deployment must look elsewhere"},{"rank":2,"product":"Qwen3","reason":"The most complete open family for application builders — Apache 2.0, sizes from sub-1B to 235B MoE sharing one behavior profile, first-rate multilingual coverage, strong function calling, and the deepest fine-tuning/quantization ecosystem, so one stack scales from edge to server; near-tie with DeepSeek, losing only on peak reasoning quality","fix":"Top-end reasoning and agentic performance still trails DeepSeek and Kimi at the frontier, so it is not the pick when maximum single-model capability matters more than deployment flexibility"},{"rank":3,"product":"Kimi K2","reason":"The strongest open-weight agentic model — its 1T-parameter MoE (32B active) leads open models on tool-calling, multi-step agent tasks, and coding benchmarks, making it the best drop-in for agent frameworks where reliability of function calls is the bottleneck; modified-MIT license is permissive in practice","fix":"At 1T parameters it is effectively hosted-inference-only for almost everyone, and the provider ecosystem is thinner than DeepSeek's or Qwen's, so latency and availability vary"},{"rank":4,"product":"gpt-oss-120b","reason":"Apache 2.0 with the best capability-per-GPU in the open field — MXFP4 quantization lets it run on a single 80GB GPU with strong reasoning and adjustable effort levels, making it the most realistic self-host option for teams that must keep data on their own hardware","fix":"Noticeably weaker world knowledge and higher hallucination rates than same-tier peers, and its safety-tuned refusals frustrate some application domains — it is not the pick for knowledge-heavy consumer products"},{"rank":5,"product":"GLM-4.6","reason":"The best open model specifically for coding-agent applications — near-Claude-Sonnet coding performance, strong long-context tool use in harnesses like Claude Code and Cline, MIT license, and cheap hosted inference; ranked on the assumption the practitioner is building developer-facing or code-generating products","fix":"Narrower general-purpose strength — outside coding and agentic loops its writing, multilingual, and knowledge quality trail the models above, so it is a specialist pick rather than a default"}],"Gemini":[{"rank":1,"product":"DeepSeek-V4-Pro","reason":"Offers frontier-level intelligence for coding, reasoning, and agentic workflows with a dual Thinking/Non-Thinking mode under a permissive MIT license, while dramatically reducing KV cache usage.","fix":"Its massive 1.6T total parameter size makes local self-hosting extremely resource-intensive, requiring high-end multi-GPU cluster infrastructure."},{"rank":2,"product":"Llama 4 Maverick","reason":"Meta's flagship 400B MoE model offers excellent generalist reasoning and coding capabilities with a 1M token context, highly optimized for fast inference with only 17B active parameters.","fix":"The Llama Community License requires a separate agreement for entities with over 700M monthly active users, creating compliance hurdles for large enterprise adopters."},{"rank":3,"product":"Qwen3-Coder-480B-A35B","reason":"Extremely strong specialized performance in multi-turn software engineering tasks and coding agents, backed by a fully permissive Apache 2.0 license for unrestricted commercial use.","fix":"Lacks the native dual reasoning-step thinking configurations of newer 2026 models, leading to potential gaps in highly abstract logical reasoning tasks."},{"rank":4,"product":"GLM-5.2","reason":"A 744B MoE released under a permissive MIT license in mid-2026, highly optimized for long-horizon agentic workflows, multi-step tool-use, and 1M-token context operations.","fix":"Requires significant memory for self-hosting and suffers from a less mature community library ecosystem outside Chinese developer circles."},{"rank":5,"product":"Gemma 4","reason":"Google's lightweight unified multimodal open-weights model supporting text, image, video, and audio on Apache 2.0, running efficiently on consumer-grade workstation GPUs.","fix":"Its 12B parameter capacity inherently limits its deep logic reasoning and broad world-knowledge compared to massive 100B+ MoE scale models."}],"Grok":[{"rank":1,"product":"GLM-5.2","reason":"Leads open-weight models on key agentic coding and long-horizon benchmarks like SWE-Bench Pro and Terminal-Bench (strong real-world software engineering performance); large MoE with excellent reasoning (e.g., top GPQA); MIT license enables broad use for building production apps.","fix":"High resource demands for full inference (multi-GPU needed for best performance); less emphasis on multimodality."},{"rank":2,"product":"Kimi K2.5","reason":"Native multimodal agentic capabilities with strong vision+coding, agent swarms, and frontier-competitive benchmarks (high SWE-Bench Verified, GPQA); efficient MoE (1T total/32B active); open weights under Modified MIT for flexible app integration.","fix":"Newer multimodal focus may require more tuning for pure text-only pipelines; inference scale still substantial."},{"rank":3,"product":"DeepSeek V4","reason":"Exceptional efficiency (MoE with low active params) and 1M+ context at competitive cost/performance; strong on math/reasoning/agentic coding; MIT license and proven self-hosting value for scalable app backends.","fix":"Slightly trails leaders on some specialized long-horizon coding evals; ecosystem less mature than Meta's."},{"rank":4,"product":"Qwen3 / Qwen3.6","reason":"Versatile family with strong coding/reasoning variants, hybrid modes, Apache 2.0 license, and excellent multilingual/long-context support; practical for diverse app development with good efficiency across sizes.","fix":"Performance can vary by specific variant; some larger ones less optimized for single-GPU local runs."},{"rank":5,"product":"Llama 4","reason":"Robust ecosystem, tooling, and community support for integration; native multimodal MoE with solid general performance and commercial-friendly license (up to usage limits); reliable for production apps.","fix":"Not always the absolute benchmark leader in specialized coding/agentic tasks compared to Chinese frontier open models."}]},"missedByModel":{"ChatGPT":[{"product":"DeepSeek-V4-Flash","reason":"excellent speed, context length, and cost, but materially weaker than V4-Pro and less locally practical than Qwen3.6-35B-A3B"},{"product":"Llama 4 Maverick","reason":"broad ecosystem and multimodality, but its capability-to-compute value no longer earns a top-five place"}],"Claude":[{"product":"Llama 4","reason":"Maverick/Scout underdelivered against 2025-26 Chinese open releases, and the Llama license's acceptable-use and branding terms make it less open than Apache/MIT rivals"},{"product":"MiniMax M2","reason":"excellent efficiency and agentic scores but a younger ecosystem and less proven fine-tuning/tooling support than the top five"}],"Gemini":[{"product":"Phi-4-mini","reason":"its small 3.8B size and lack of multimodal support limit its value for general-purpose app development compared to Gemma 4"},{"product":"Llama 4 Scout","reason":"its 10M context window is impressive, but hosting a 109B total parameter model for 17B active performance is less cost-efficient than utilizing Maverick or V4-Pro"}],"Grok":[{"product":"MiniMax M2.5/M3","reason":"strong multimodal/coding contender but narrower lead and less ecosystem depth"}]}}