Best open-weight LLM
4 models · updated 2026-07-15
The verdict
GLM-5.2 leads — 2 of 4 models rank GLM-5.2 the top pick.
Not unanimous: Claude picks DeepSeek-V3.2; Gemini picks DeepSeek-V4-Pro.
Your vendor missing? Check any brand →
Combined ranking
- 1GPT #1Claude —Gemini #4Grok #1
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.
+ model takes & fixes− hide details
GPT 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.
Grok 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.
Gemini 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.
Where it falls shortper GPT Its roughly 753B-weight footprint makes practical self-hosting a multi-GPU or datacenter undertaking.
per Gemini Requires significant memory for self-hosting and suffers from a less mature community library ecosystem outside Chinese developer circles.
per Grok High resource demands for full inference (multi-GPU needed for best performance); less emphasis on multimodality.
- 2GPT #2Claude —Gemini #1Grok —
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.
+ model takes & fixes− hide details
Gemini 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.
GPT Near-tied with GLM-5.2, with exceptional reasoning, coding, long-context processing, agentic execution, a permissive license, and unusually strong API value.
Where it falls shortper GPT The enormous 1.6T-parameter model is unrealistic for typical practitioners to self-host.
per Gemini Its massive 1.6T total parameter size makes local self-hosting extremely resource-intensive, requiring high-end multi-GPU cluster infrastructure.
- 3GPT —Claude #1Gemini —Grok —
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
+ model takes & fixes− hide details
Claude 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
Where it falls shortper Claude 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
- 4GPT —Claude —Gemini —Grok #2
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.
+ model takes & fixes− hide details
Grok 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.
Where it falls shortper Grok Newer multimodal focus may require more tuning for pure text-only pipelines; inference scale still substantial.
- 5GPT —Claude —Gemini #2Grok —
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.
+ model takes & fixes− hide details
Gemini 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.
Where it falls shortper Gemini The Llama Community License requires a separate agreement for entities with over 700M monthly active users, creating compliance hurdles for large enterprise adopters.
- 6GPT —Claude #2Gemini —Grok —
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
+ model takes & fixes− hide details
Claude 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
Where it falls shortper Claude 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
- 7GPT #5Claude #4Gemini —Grok —
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
+ model takes & fixes− hide details
Claude 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
GPT Mature open tooling, Apache 2.0 licensing, strong reasoning and function calling, and single-80GB-GPU operation make it a dependable customizable production choice.
Where it falls shortper GPT Its text-only capabilities and older performance ceiling now trail newer open-weight leaders.
per Claude 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
- 8GPT —Claude —Gemini —Grok #3
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.
+ model takes & fixes− hide details
Grok 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.
Where it falls shortper Grok Slightly trails leaders on some specialized long-horizon coding evals; ecosystem less mature than Meta's.
- 9GPT —Claude #3Gemini —Grok —
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
+ model takes & fixes− hide details
Claude 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
Where it falls shortper Claude 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
- 10GPT #3Claude —Gemini —Grok —
Particularly strong for repository-scale coding and autonomous software-engineering workflows, making it an excellent foundation for coding agents and application builders.
+ model takes & fixes− hide details
GPT Particularly strong for repository-scale coding and autonomous software-engineering workflows, making it an excellent foundation for coding agents and application builders.
Where it falls shortper GPT Its coding specialization makes it less compelling than the top generalists for broad knowledge, writing, or mixed enterprise workloads.
- 11GPT —Claude —Gemini #3Grok —
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.
+ model takes & fixes− hide details
Gemini 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.
Where it falls shortper Gemini Lacks the native dual reasoning-step thinking configurations of newer 2026 models, leading to potential gaps in highly abstract logical reasoning tasks.
- 12GPT —Claude —Gemini —Grok #4
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.
+ model takes & fixes− hide details
Grok 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.
Where it falls shortper Grok Performance can vary by specific variant; some larger ones less optimized for single-GPU local runs.
- 13GPT #4Claude —Gemini —Grok —
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.
+ model takes & fixes− hide details
GPT 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.
Where it falls shortper GPT It cannot match the frontier models above on the hardest long-horizon reasoning and agentic tasks.
- 14GPT —Claude —Gemini #5Grok —
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.
+ model takes & fixes− hide details
Gemini 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.
Where it falls shortper Gemini Its 12B parameter capacity inherently limits its deep logic reasoning and broad world-knowledge compared to massive 100B+ MoE scale models.
- 15GPT —Claude #5Gemini —Grok —
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
+ model takes & fixes− hide details
Claude 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
Where it falls shortper Claude 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
- 16GPT —Claude —Gemini —Grok #5
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.
+ model takes & fixes− hide details
Grok 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.
Where it falls shortper Grok Not always the absolute benchmark leader in specialized coding/agentic tasks compared to Chinese frontier open models.
Just missed the top 5
GPT DeepSeek-V4-Flash — excellent speed, context length, and cost, but materially weaker than V4-Pro and less locally practical than Qwen3.6-35B-A3B · Llama 4 Maverick — broad ecosystem and multimodality, but its capability-to-compute value no longer earns a top-five place
Claude Llama 4 — 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 · MiniMax M2 — excellent efficiency and agentic scores but a younger ecosystem and less proven fine-tuning/tooling support than the top five
Gemini Phi-4-mini — its small 3.8B size and lack of multimodal support limit its value for general-purpose app development compared to Gemma 4 · Llama 4 Scout — 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 MiniMax M2.5/M3 — strong multimodal/coding contender but narrower lead and less ecosystem depth
By model
ChatGPT
- 1.GLM-5.2
- 2.DeepSeek-V4-Pro
- 3.Kimi K2.7 Code
- 4.Qwen3.6-35B-A3B
- 5.gpt-oss-120b
Claude
- 1.DeepSeek-V3.2
- 2.Qwen3
- 3.Kimi K2
- 4.gpt-oss-120b
- 5.GLM-4.6
Gemini
- 1.DeepSeek-V4-Pro
- 2.Llama 4 Maverick
- 3.Qwen3-Coder-480B-A35B
- 4.GLM-5.2
- 5.Gemma 4
Grok
- 1.GLM-5.2
- 2.Kimi K2.5
- 3.DeepSeek V4
- 4.Qwen3 / Qwen3.6
- 5.Llama 4
This ranking moves
We re-poll all four models weekly. Get one short email when a #1 flips.
Tracked by ModelsAgree · rank 1 = 5 pts … rank 5 = 1 pt · re-polled weekly