The Best Open-Source and Open-Weight LLMs in 2026

Key Takeaways

• Most models labeled "open" are open-weight, not fully open-source; under OSAID 1.0, open-source AI requires model parameters, complete training and inference code, and sufficiently detailed information about the training data under OSI-approved terms.

• GLM-5.3 Flash is the current GLM entry in this guide, with a 1,000,000-token context window, 320B total parameters, 18B active parameters, native multimodal input, and MIT-licensed weights.

• DeepSeek-V4-Pro activates 49B of its 1.6T parameters per token, which affects compute; self-hosting still requires storing or loading the full weight set unless the deployment uses offloading or another distributed strategy.

• Mistral Small 4 is an Apache 2.0-licensed candidate for teams evaluating a unified instruct, reasoning, coding, and multimodal model; current hosted pricing should be checked on the selected provider route.

• Kimi K3 is a 2.8T-parameter multimodal model that accepts text, image, and video inputs; it is not the only multimodal model in this guide.

• MiniMax M3 supports text, image, and video inputs, making it a candidate for mixed-media workflows; Kimi K3 and other entries also provide multimodal capabilities.

• Llama 4 is supported across many inference, quantization, and fine-tuning tools. Ecosystem breadth does not by itself establish a performance ranking, so teams should test the relevant variant and workload.

Open-Source vs. Open-Weight vs. Open-Access: What the Labels Actually Mean

These three labels describe very different legal and technical realities, not interchangeable marketing terms. Open-weight means the trained model weights are publicly downloadable, but the license may restrict commercial use, fine-tuning redistribution, or deployment at scale. Under OSAID 1.0, open-source AI requires model parameters, complete code used to train and run the system, and sufficiently detailed information about the training data under OSI-approved terms.

Apache 2.0 and MIT are permissive licenses, but they still impose conditions such as preserving copyright and license notices; Apache 2.0 also contains patent provisions. Most models on this list are open-weight, not fully open-source by the OSI definition.

The distinction matters before adoption because it determines your legal freedoms. There are 3 categories to distinguish:

• Open-source: weights, source code, and other required components released under an OSI-approved license; commercial use and redistribution remain subject to that license's conditions

• Open-weight: weights are downloadable; license terms vary and frequently prohibit use above a defined user or revenue threshold, or require a separate commercial agreement

• Open-access: the model is accessible only through a hosted API or interface; weights are never released; the operator controls all deployment decisions

A model labeled "open" by its developer is not automatically open-source. Llama models, for example, carry a custom Meta community license that imposes usage restrictions, placing them in the open-weight category rather than the open-source category (source).

Verify the exact license file in a model's repository before deploying it in a production environment, if your project involves commercial distribution or user-facing products.

How We Chose and Evaluated These Models

Every model on this list passed 5 mandatory filters. These filters cover openly available weights and a license permitting at least research use. They also require documented benchmark results on at least 1 public reasoning or coding leaderboard, a stated context window, and active maintenance as of 2026.

There are 5 evaluation dimensions we applied to each candidate:

• License and commercial terms — we read the exact license file in each model's repository and classified it as permissive, restricted, or research-only.

• Weight availability — we confirmed the weights are downloadable or accessible via a public API endpoint, not gated behind a waitlist.

• Benchmark performance — we cross-referenced scores from public composite intelligence indices such as the Open LLM Leaderboard and LiveBench, covering reasoning, coding, math, and multilingual tasks.

• Context window — we recorded the officially documented maximum context length for each model.

• Multimodality — we noted whether the model natively accepts image, audio, or video inputs alongside text.

This guide is based on official model cards, repositories, license files, release notes, and provider documentation. We compared documented architecture, context limits, modalities, licensing, and published benchmark results.

We did not run a controlled benchmark across all models under one harness, quantization level, hardware setup, or prompt set. Vendor-reported results are treated as directional, and every recommendation should be validated on the intended workload.

The Best Open-Source and Open-Weight LLMs in 2026, Ranked

This guide covers 11 current open-weight models across coding, chat, RAG, multimodal, and agent workflows. The order is an editorial shortlist based on documented breadth, not a controlled cross-model benchmark.

Each profile covers the developer, license, context window, multimodal support, and documented fit. Published results from different vendors may use different harnesses and should not be treated as directly comparable.

1. GLM-5.3 Flash

GLM-5.3 Flash is developed by Z.ai and released under the MIT License. Its official materials describe 320B total parameters, 18B active parameters, a 1,000,000-token context window, and native support for text, image, video, and file inputs with text output.

Z.ai positions GLM-5.3 Flash for efficient coding, agent, long-context, and multimodal workloads. Its MIT-licensed weights support self-hosting, but production fit still depends on task quality, serving requirements, hardware capacity, and compliance with the license terms.

2. DeepSeek-V4-Pro

DeepSeek-V4-Pro is developed by DeepSeek and uses a mixture-of-experts architecture with 1.6T total parameters and 49B active parameters. The context window reaches 1,000,000 tokens, and the model is text-only. Benchmark guides published in 2026 cite competitive SWE-Bench Verified and reasoning scores (source).

DeepSeek-V4-Pro is a candidate for long-context reasoning and large-codebase workloads. The 49B active-parameter figure affects per-token compute, but the complete 1.6T-parameter weight set still drives storage, loading, memory, and distributed-serving requirements.

3. Kimi K3

Kimi K3 is developed by Moonshot AI under the Kimi K3 License and uses a 2.8T-parameter mixture-of-experts configuration. Its official model materials describe an approximately 1,000,000-token context window and native text, image, and video input.

Kimi K3 may fit long-context, multimodal research and agent workflows. Its custom license requires review before redistribution or commercial deployment, and teams should test quality and serving cost on representative text, image, and video inputs.

4. Kimi K2.7 Code

Kimi K2.7 Code is Moonshot AI's code-focused open-weight model. Its official repository documents a 1T-parameter model, a 262,144-token context window, multimodal image input, and the repository's modified MIT license; teams should review that license before commercial redistribution.

Kimi K2.7 Code is designed for long-horizon coding agents, IDE copilots, and software-engineering workflows. Teams should compare it with other coding-focused models under the same repository, tool, and evaluation harness before selecting it.

5. MiniMax M3

MiniMax M3 is developed by MiniMax under the MiniMax Community License, with 428B total parameters and 23B active parameters. The context window reaches 1,000,000 tokens (with 512,000 tokens guaranteed at standard tier), and the model accepts text, images, and video. API pricing starts at $0.30 per million input tokens and $1.20 per million output tokens.

MiniMax M3 supports long-context text, image, and video workflows. It is one candidate for mixed-media agents, but its quality, license terms, provider availability, and total cost should be validated for the intended deployment.

6. Qwen3.5 (family)

Qwen3.5 is an Alibaba Cloud model family with variants up to 397B total parameters and 17B active parameters. The official 397B model card documents a 262,144-token native context window, extension to 1,010,000 tokens with YaRN, Apache 2.0 licensing, and native multimodal support.

The Qwen3.5 family spans multiple deployment sizes and may fit multilingual chat, structured extraction, coding, and multimodal workloads. Developers can browse the Qwen models available on 开桥 before selecting and testing a model size, context configuration, and serving stack for the intended task.

7. Llama 4

Llama 4 is developed by Meta under the Llama Community License and is available as free, self-hosted open weights. The family includes variants such as Llama 4 Scout and Llama 4 Maverick, all natively multimodal (source). Context windows are reported in the multi-million token range for select variants. The Llama ecosystem is broadly supported on Hugging Face with extensive tooling.

Llama 4 benefits from broad support across common inference, quantization, and fine-tuning tools. Teams already using the Llama ecosystem may reduce integration work, but should still review the Meta license and test the selected variant on their workload.

8. Gemma 4

Gemma 4 is developed by Google DeepMind under the Apache 2.0 license, with a family range spanning approximately 2.3B to 31B parameters. The context window reaches 256K tokens, and the model supports multimodal input.

Gemma 4 is an Apache 2.0-licensed family with compact variants that may fit on-device or privacy-sensitive deployments. Actual device fit depends on model size, quantization, runtime overhead, and available memory.

9. Mistral Small 4

Mistral Small 4 is developed by Mistral AI under the Apache 2.0 license, with 119B total parameters and 6.5B active parameters. The model accepts multimodal input and is positioned as a unified instruct, reasoning, and coding model (source). API pricing is $0.15 per million input tokens and $0.60 per million output tokens.

Mistral Small 4 combines Apache 2.0 licensing with a unified instruct, reasoning, coding, and multimodal design. Teams should compare current provider pricing and evaluate quality, latency, and memory requirements on their own workload.

10. NVIDIA Nemotron 3 Super 120B

NVIDIA Nemotron 3 Super 120B is developed by NVIDIA under the NVIDIA Nemotron Open Model License, with 120B total parameters and 12B active parameters (source). The architecture combines Mamba-2, mixture-of-experts, and standard attention layers. The context window reaches 1,000,000 tokens.

Nemotron 3 Super 120B is optimized for NVIDIA infrastructure and long-context agentic and RAG workloads. Teams should review its license and deployment requirements, then test throughput and retrieval quality on their target NVIDIA stack.

11. gpt-oss-120b

gpt-oss-120b is developed by OpenAI under the Apache 2.0 license. The model name is rounded: OpenAI's official model card lists 116.8B total parameters and a context window of 131,072 tokens (source). The model is text-only and is governed by both the Apache 2.0 license and OpenAI's gpt-oss usage policy.

gpt-oss-120b may fit teams using OpenAI-compatible tooling that want a self-hostable reasoning model. Its 131,072-token context window is shorter than several large models in this guide, so document-heavy RAG may require chunking or retrieval.

Open-Weight LLM Comparison Table: License, Context, Size & Best Fit

Scan the License and Max Context Window columns first, then match the Best Fit column to your workload.

Benchmark Performance: Reasoning, Coding, Math and Multilingual

Published model cards report results across reasoning, coding, math, agent, and multilingual evaluations. Because vendors use different harnesses, settings, and sample counts, these figures indicate areas to test rather than a single defensible cross-model ranking.

Reasoning

Kimi K3, DeepSeek-V4-Pro, and GLM-5.3 Flash each publish reasoning results, but those results were not produced under one shared evaluation setup. Compare them with the same prompts, tools, reasoning budget, and scoring method.

Coding and Agent Tasks

Kimi K2.7 Code, GLM-5.3 Flash, DeepSeek-V4-Pro, Mistral Small 4, and gpt-oss-120b publish coding or agent-related results. The scores are not directly comparable unless the same harness, repository set, tools, and reasoning settings are used.

Math

DeepSeek-V4-Pro, Kimi K3, GLM-5.3 Flash, and Qwen3.5 publish math or reasoning results. Treat vendor-reported scores as directional until the models are evaluated under one shared setup.

Multilingual Coverage

Qwen3.5 and MiniMax M3 document multilingual capabilities. Language coverage and output quality vary by language and task, so production teams should test their target locales directly.

Reading Composite Indices Without Over-Trusting a Single Score

Composite benchmark indices aggregate scores across task types. A model that dominates one category can rank highly overall while showing limited capability in another. Evaluate each model against the specific task distribution of a real project.

A leading reasoning composite score does not guarantee strength on domain-specific coding or low-resource language generation. Cross-referencing at least 3 independent benchmark sources before selecting a model for production reduces the risk of misalignment between benchmark rank and real-world output quality.

Licensing and Commercial-Use Permissions Explained

Apache 2.0 and MIT are permissive licenses that generally allow commercial use, modification, and redistribution, subject to their terms. Required notices, Apache 2.0 patent provisions, trademarks, model-specific policies, and applicable law still need review. Qwen3.5, Gemma 4, Mistral Small 4, gpt-oss-120b, and DeepSeek-V4-Pro publish weights under one of these frameworks.

Custom and community licenses occupy a different legal category. Llama 4 carries Meta's custom license, which imposes usage restrictions on services that exceed defined monthly active user thresholds (source). Redistributing Llama 4 weights requires retaining Meta's attribution notice and prohibits using the model to train competing foundation models. These restrictions do not disqualify Llama 4 for most enterprise deployments, but they require explicit legal review before shipping.

Fine-tuning and derivative-work rights vary by license. Apache 2.0 permits derivative works subject to its notice and patent terms, while custom licenses may add restrictions that affect a team's IP strategy. Review the exact license version before investing in a fine-tuning pipeline.

There are 5 due-diligence steps to complete before deploying any open-weight model commercially:

1. Read the full license text on the model's official repository page, not a summary.

2. Identify whether a monthly active user or revenue threshold triggers additional obligations.

3. Confirm redistribution terms if the model is embedded in a shipped product or API.

4. Verify fine-tuning and derivative-work clauses against the team's IP ownership requirements.

5. Document the license version pinned at deployment, because license terms can change between model releases.

Hardware and VRAM Requirements: A Self-Host Reality Check

Self-hosting requirements depend on total parameters, precision, KV cache, runtime overhead, parallelism, and whether weights are kept in GPU memory or offloaded. Compact quantized variants may fit a consumer GPU; large MoE models generally require multi-GPU or distributed serving.

MoE architecture reduces the parameters used for compute on each token, but it does not reduce the complete weight set that must be stored or made available to the serving system. Active-parameter counts should not be used as a proxy for model-weight memory.

Quantization reduces weight storage, but it does not remove KV-cache and runtime overhead. As a weights-only lower bound, 4-bit storage is roughly 0.5 bytes per parameter: about 15.5GB for 31B parameters, 59.5GB for 119B, 160GB for 320B, 800GB for 1.6T, and 1.4TB for 2.8T. BF16 weights require roughly four times those amounts.

There are 3 practical deployment tiers based on available hardware:

• Consumer single GPU (8–24GB VRAM): compact quantized variants may fit, depending on model size, KV cache, context length, runtime, and offloading; a 119B model does not fit this tier on weights alone at 4-bit

• Workstation or high-memory GPU (48–80GB VRAM): some larger quantized models may fit, but a 119B model already needs about 59.5GB for 4-bit weights before runtime and KV-cache overhead

• Multi-GPU or distributed server: size capacity for the selected model and precision; 160GB is only the weights-only 4-bit lower bound for a 320B model, while 1.6T and 2.8T models require far more memory

On-device deployment on laptops or phones remains limited to the smallest quantized variants. Inference speed and output quality degrade noticeably compared to GPU-hosted runs. Teams without server-class hardware can use the 开桥 model catalog to access supported models through a unified API without managing local VRAM constraints.

How to Run or Access These Models: Local, Self-Host, or Hosted API

Open-source and open-weight LLMs reach production through 3 distinct deployment paths: local runners for experimentation, self-hosted GPU infrastructure for control, and hosted inference APIs for zero-ops access.

Local Runners: Ollama and vLLM

Ollama packages model weights with a one-command runtime, letting a developer pull and run quantized models on a laptop CPU or consumer GPU within minutes. vLLM targets higher-throughput local or on-premises use, applying PagedAttention to maximize tokens-per-second on NVIDIA hardware. Both tools suit experimentation and prototyping. On-device runs with quantized weights produce noticeably lower output quality than full-precision GPU-hosted inference — local runners are the right entry point, not the production endpoint.

Self-Hosting on Your Own GPUs

Self-hosting gives teams full data residency and deterministic latency, at the cost of provisioning, VRAM budgeting, and ongoing maintenance. Teams that handle regulated data or require air-gapped environments choose this path. A model that exceeds available VRAM cannot run without quantization trade-offs, so the deployment-tier guidance above applies directly.

Hosted Inference via a Unified API Router

Teams that want access to multiple supported models without managing GPU infrastructure can use a unified API. 开桥 provides one integration for models currently listed in its catalog and routes supported requests through official providers or authorized cloud partners. Availability, provider routes, and pricing should be checked on the live 开桥 model pages.

Cost Comparison: Self-Host GPU vs. Hosted API Access

The right choice between self-hosting and hosted API access depends on inference volume and workload predictability. Self-hosting favors high, steady inference volume where amortized GPU cost falls below per-token API pricing. Hosted API access wins for variable or early-stage workloads where capital expenditure and operational overhead are not justified.

GPU infrastructure carries 3 distinct cost layers: hardware acquisition, power and cooling, and engineering time for deployment, monitoring, and model updates. A team running a large open-weight model at low utilization pays the full fixed cost regardless of actual request volume. Utilization below a break-even threshold makes every token more expensive than buying it from a hosted provider.

Per-token pricing across the roster models spans a wide range. Frontier-class models with strong reasoning and multimodal capabilities sit at the premium end. Smaller, Apache-licensed models like Mistral Small 4 and Gemma 4 occupy the cost-efficient end. The break-even point shifts depending on which model a team targets: a smaller model requires less GPU memory and cheaper hardware, lowering the volume threshold at which self-hosting becomes economical.

Hosted API access removes the utilization risk entirely. A team pays only for tokens consumed, with no idle GPU cost during low-traffic periods. This structure fits prototyping, seasonal workloads, and multi-model experimentation — exactly the pattern where routing infrastructure adds measurable value.

开桥 provides unified access to supported models through official providers or authorized cloud partners. Teams can compare current routes and pay per use without provisioning model-serving infrastructure for every supported model.

开桥 also publishes AI Model Insurance for eligible scenarios. Teams should review current eligibility, coverage, model availability, provider routes, and policy terms on 开桥 before deployment.

Choosing the Right Open-Weight Model by Use Case

The right open-weight model depends on the primary workload, not benchmark rank alone. Coding agents, general chat, multimodal pipelines, on-device deployment, and enterprise NVIDIA stacks each demand a different model. Choosing well means matching architecture and tooling to the actual task at hand.

Coding Agents

Kimi K2.7 Code and GLM-5.3 Flash are candidates for coding-agent, IDE-copilot, and DevOps-assistant workflows. Compare task completion, tool-call validity, repository-scale accuracy, latency, and serving cost under the same harness.

General Chat and Instruction Following

Qwen3.5 and Mistral Small 4 are candidates for general chat and instruction-following workloads. Test target languages, latency, fine-tuning support, and deployment cost before choosing a variant.

Multimodal Agents

MiniMax M3 and Kimi K3 are candidates for document, image, video, and long-running multimodal-agent workflows. Input support, license terms, provider availability, and task quality should be verified for the selected deployment.

On-Device and Privacy-Sensitive Deployments

Gemma 4 may fit on-device or privacy-sensitive deployments that need Apache 2.0-licensed weights. Confirm that the chosen variant, quantization, context length, and runtime fit the target device and quality requirements.

Enterprise and NVIDIA Stacks

Nemotron 3 Super may fit teams already operating NVIDIA infrastructure and tooling. Confirm TensorRT-LLM, NeMo, licensing, hardware, and support requirements for the intended deployment.

Frequently Asked Questions

What is the difference between an open-source LLM and an open-weight LLM?

Under OSAID 1.0, an open-source AI system must provide model parameters, complete training and inference code, and sufficiently detailed information about the training data under OSI-approved terms. An open-weight model releases trained weights but may not provide all of those components. Many models marketed as "open-source" in 2026 are therefore more accurately described as open-weight.

Which open-weight LLM is the best overall in 2026?

There is no single defensible winner across all workloads in this guide. GLM-5.3 Flash, DeepSeek-V4-Pro, Kimi K3, and the other listed models differ in modality, context, license, serving requirements, and published evaluations; test the shortlist on the intended workload.

Can I use these open-weight models commercially for free?

Commercial-use rights vary by model and release. Llama 4 uses Meta's community license, while Mistral Small 4, Gemma 4, Qwen3.5, gpt-oss-120b, and DeepSeek-V4-Pro publish weights under Apache 2.0 or MIT. Review the exact repository license, required notices, model-specific policies, and any scale thresholds before deployment.

How much VRAM do I need to run a frontier open-weight LLM locally?

Memory requirements for frontier open-weight models scale steeply with precision. Running an 80-billion-parameter model at full FP32 precision demands approximately 297 GB of memory just to load its weights (source). Quantized versions at 4-bit precision reduce that requirement to approximately 37 GB, although additional memory is needed for the KV cache and inference overhead.

Is it cheaper to self-host an open-weight model or use a hosted API?

The lower-cost option depends on utilization, model size, precision, hardware, energy, staffing, cache behavior, provider pricing, and required availability. Compare workload-level total cost rather than assuming either self-hosting or hosted access is cheaper.

Which open-weight model is best for building coding agents?

Among the models covered in this guide, Kimi K2.7 Code, GLM-5.3 Flash, DeepSeek-V4-Pro, Mistral Small 4, and gpt-oss-120b are candidates for coding agents. Compare them under the same repository tasks, tools, reasoning settings, and completion criteria.

How can I try and switch between multiple open-weight LLMs without provisioning GPUs?

A unified API can provide access to supported hosted models without requiring a separate integration for every provider. 开桥 lists a subset of the models discussed in this article; verify each exact model and route in the current catalog before testing or switching.