The Best Open-Source LLMs in 2026: 8 Open-Weight Models Ranked

Start with GLM-5.3 and Mistral Large 4 for a coding-agent evaluation, DeepSeek V4.1 Flash-1017 when MIT-licensed weights are the priority, and gpt-oss-120b if your deployment budget is one 80GB GPU. Large 4 now brings downloadable text-and-image weights to the shortlist, but its custom license and roughly trillion-parameter size make it a different deployment decision from Mistral Small 4.

The verdict

If you want an open model for long coding jobs, evaluate GLM-5.3 first and Mistral Large 4 alongside it, subject to their current licenses. For coding and document work under MIT, start with DeepSeek V4.1 Flash-1017. If an Apache 2.0 checkpoint is a requirement, shortlist gpt-oss-120b, Mistral Small 4, or the older Qwen3-235B-A22B. Muse Spark Open belongs on the shortlist when your agent needs audio as well as visual and text input.

Since October 29, 2026, Mistral Large 4's weights have been available on Hugging Face, making it a candidate for deployment in your own cloud or on-premises. The model cards and licenses below were re-verified on October 30, 2026; the shortlist also replaces the original DeepSeek V4.1 Flash with Flash-1017 and moves the MiniMax evaluation to M3.1 Flash. Weight-release announcement, Hugging Face release.

An open-weight model is one whose trained parameters you can download and run yourself. That gives you another deployment option; it does not make serving free. These eight primary picks are ordered for this buying decision, balancing capabilities, license conditions and deployment practicality. They are an analysis of published evidence, not results from a hands-on comparison. Large 4 joins the coding shortlist on Mistral's published evidence; its position is not a claim that it beats every model on every task.

Context is the amount of material a model can handle in a request. It is not a promise that every host exposes that limit, or that every fact in a long document will be retrieved correctly. Check the exact checkpoint, host and serving configuration before sizing your application around it.

Why open weights stopped being a compromise

You can now choose downloadable models for coding, tool use and visual-document workflows without treating them as a separate class of toy systems. The useful question is whether a particular model completes your tasks within your latency, hardware and license constraints.

Benchmarks help narrow that evaluation, but the surrounding software matters. A harness is the program that gives the model tools, prompts and retry rules. Mistral's released card reports 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 in that comparison, while using 47% fewer tokens per task. That earns Large 4 a coding evaluation beside GLM; it does not establish the fastest deployment or an overall intelligence winner. Mistral's own card says the leading Chinese open-weight models score higher on the aggregate intelligence index. Mistral Large 4 model card.

Open weight and open source are different claims. Download access tells you that you can obtain the parameters. A permissive weight license tells you what you may do with them. The Open Source Initiative's definition goes further, requiring appropriate access to code and training-data information as well as parameters. An MIT or Apache label on weights alone does not establish that the whole AI system meets that definition. Open Source AI Definition.

GLM-5.3 is the coding model to evaluate first

GLM-5.3, from Z.ai, is the first pick here for a team evaluating long coding and tool workflows. Its BF16 model page lists 753B parameters; its current documentation specifies text-only input, a 1M-token context window and always-on reasoning. That makes it a candidate for repository work, but not a standalone model for reading screenshots. A large context window can help keep code and instructions together; it does not eliminate the need to retrieve the right files. GLM-5.3 model card, documentation.

For a coding agent, begin with the maker's recommended reasoning setting and compare completed patches, failed tool calls and elapsed time against your existing model. Z.ai's card lists deployment support including SGLang, vLLM and Transformers. These are inference engines, the software that loads the weights and serves requests. The choice of engine and quantization, which stores weights at lower precision, is part of your deployment specification. GLM-5.3 model card.

GLM-5.3 is not MIT: its license contains a security-review requirement for defined model-as-a-service businesses whose licensee-and-affiliate aggregate revenue exceeds $10 billion over any consecutive 12 months. If standard permissive terms are your deciding constraint, evaluate DeepSeek or an Apache checkpoint instead of carrying a previous GLM release's license label onto this one. GLM-5.3 license.

Mistral Large 4 joins the coding and document shortlist

Mistral Large 4 is the Mistral checkpoint to evaluate for a large coding agent that also reads images, charts and technical documents. Its Hugging Face card specifies approximately 1T total parameters, 49B active parameters per token, or 52B including embeddings and output layers, and a 1M-token context. It accepts text and images and produces text. The 49B and 52B figures describe different counting boundaries, not different models. Mistral Large 4 card.

For an internal engineering agent that reads a diagram, changes code and returns structured results, Large 4 combines those inputs with function calling and strict JSON-schema output. Its reasoning effort can be set to low, medium or high, with high the default. That is a reason to compare it with GLM on your full workflow, especially when visual input matters. The published coding-index tie earns it a place beside GLM, not an automatic migration.

The released weights use the Mistral Ecosystem Open-Weight License, a modified MIT license. If you or an affiliate operates the defined model-as-a-service business, commercial use requires a separate Mistral license; this clause has no revenue threshold. The license exempts defined internal use and use through Mistral's official products. Its model-as-a-service definition excludes capabilities embedded solely within specific end-user features or harnesses and mere relaying to models hosted elsewhere. An internal on-premises assistant and a general inference endpoint sold to customers therefore need different license assessments. Large 4 license.

The official releases include FP8 and NVFP4 weights. These are lower-precision storage formats, not smaller architectures. Mistral's NVFP4 card targets NVIDIA Blackwell GPUs and documents vLLM and SGLang serving. Treat this as a cluster-sizing exercise: the active parameter count does not tell you how much memory the full checkpoint and concurrent requests need. NVFP4 card and deployment recipe.

Keep Mistral Small 4 when Apache 2.0 or a smaller model is the deciding constraint. Small 4 remains 119B total / 6.5B active parameters with 256K context, text-and-image input, function calling and JSON output. Large 4 does not supersede that license-and-size combination. For scanned-form extraction, evaluate Small 4 before committing to Large 4's footprint. Its per-request reasoning settings also remain useful for separating ordinary extraction from harder interpretation. Mistral Small 4 card.

Neither the name Small nor the vendor's location establishes a laptop fit or European data residency. Choose the serving location, retention policy and tool integrations that meet your requirements. This is a model choice; the deployment is a separate decision.

DeepSeek V4.1 Flash-1017 is the MIT-licensed starting point

DeepSeek V4.1 Flash-1017 is the checkpoint to evaluate for coding, extraction and document agents when you want MIT-licensed weights on infrastructure you control. It replaces the original V4.1 Flash recommendation: DeepSeek describes it as an update with improvements across reasoning, coding, agentic tasks and visual benchmarks. Its current model page lists 763B parameters, text and image input, and a 1M-token context. Flash-1017 model card.

Do not read Flash as a small-hardware promise. The previous card's 552B figure counted backbone parameters, with conditional memory described separately; the current model page's 763B is its listed checkpoint size. Those are different accounting boundaries. A mixture-of-experts model activates only part of its weights for each token; the inactive weights still matter when planning storage and serving memory. Original Flash architecture, current checkpoint.

Pin the full checkpoint name in your deployment. The new card does not provide a Jinja chat template, the common message-formatting template used by many servers. It supplies reference encoding code and a protocol toolkit, while leaving inference and tool execution to your application. A compatible message encoder is part of the model integration, not an optional formatting detail. Prompt encoding and local inference.

For a founder processing incoming documents, begin with representative files and measure which cases fail. Move difficult cases to another model only when your evaluation demonstrates a benefit. A larger model is an alternative to test, not an automatic upgrade for every request.

Kimi K3 is built for long multimodal agent work

Kimi K3, from Moonshot AI, is the newer model to evaluate for coding agents that also need visual understanding or broader knowledge work. Its card specifies 2.8T total parameters, 104B active parameters and a 1,048,576-token context window, with text, image and video understanding. Kimi K3 model card.

For an internal development agent that must inspect a screenshot, change the application and check the result, a multimodal model can keep those steps within one model workflow. That is a reason to evaluate K3 alongside a text-focused coding model. It is not a reason to buy the hardware before testing the workflow through a host.

K3 uses the Kimi K3 License. Its license includes a separate-agreement clause for qualifying model-as-a-service businesses, attribution conditions and explicit exceptions for defined internal use and access through official products or certified inference partners. Read those clauses against your intended deployment. Kimi K3 license.

The trade-off is deployment size and custom licensing. A team that needs a compact, permissively licensed text model should look elsewhere in this list. For the software that gives these models their tools and execution loop, see the best AI coding agents.

Qwen3 remains a legacy permissive deployment option

Qwen3-235B-A22B, from Alibaba's Qwen team, remains a legacy choice when you specifically want an Apache 2.0 checkpoint with multilingual support. Its card lists 235B total parameters, 22B active parameters, and support for more than 100 languages and dialects. It also supports switching between thinking and non-thinking modes. Qwen3-235B-A22B card, license.

Be precise about context: this original checkpoint has 32,768 native tokens, expandable to 131,072 with YaRN, a context-extension method. It is not the same release as every later model with a similar name. If you are building multilingual support-ticket triage, test the languages you actually receive and the context configuration you plan to serve.

It is also no longer true that Max-class Qwen weights are unavailable. Qwen3.8-2.4T-A95B is a published Max-class checkpoint with 2.4T total / 95B active parameters and its own custom license. Its open checkpoint is text-only, with 262,144 native context tokens, extensible to 1,010,000. Its hosted Qwen3.8-Max service adds features beyond the released language-model checkpoint. Keep this older Qwen3 entry on your shortlist for its specific license and deployment profile, not because it is the newest Qwen flagship. Qwen3.8 card, Qwen3.8-Max license.

MiniMax M3.1 Flash is the newer long-context evaluation

MiniMax M3.1 Flash is the newer MiniMax option to evaluate when your work mixes text, images and video. Its card describes approximately 428B total parameters, 23B active parameters and up to 1,000,000 context tokens. Use the maker's architecture count here: the Hugging Face sidebar separately displays 220B, while the card's architecture section says approximately 428B. Do not treat the sidebar value as evidence that the architecture has halved in size. MiniMax M3.1 Flash card.

For a product that summarizes a recording alongside related documents, the multimodal input is the reason to evaluate it. Use representative recordings and documents, then check whether the model cites the right evidence. Context capacity alone does not tell you how reliably it will find a brief event inside a long recording.

M3.1 Flash builds on M3 with sparse attention throughout, quantization-aware training and a speculative-decoding draft head, techniques intended to reduce serving work. The original M3 is now the predecessor for this evaluation, not a disappeared checkpoint: its 428B / 23B / 1M specification and enabled, adaptive and disabled reasoning modes remain documented. M3.1 Flash instead documents low, medium, high, xhigh and max reasoning effort; do not transfer M3's controls without checking your server. M3.1 Flash changes, original M3 card.

The MiniMax Community License requires Built with MiniMax M3.1-Flash attribution for commercial use. It calls for prior written authorization when the relevant products or services generate more than $20 million in yearly revenue; otherwise it requires a one-time notice. Its prohibited-use terms also apply. Those conditions differ from Muse's model-as-a-service clause, even though both contain a revenue figure. M3.1 Flash license.

Choose it when the multimodal workflow earns its place and those terms fit. If avoiding custom license conditions is the priority, start with the Apache entries instead.

gpt-oss is the one that runs on your own GPU

gpt-oss, OpenAI's open-weight family, is the practical starting point when you need local text reasoning without a large cluster. The released MXFP4 version of gpt-oss-120b has 117B total parameters and 5.1B active parameters and fits on one 80GB GPU. The smaller gpt-oss-20b has 21B total / 3.6B active parameters and runs within 16GB of memory. Both use Apache 2.0. 120b model card, 20b model card.

This is a deployment recommendation, not a compliance guarantee. Your serving process, logs, retrieval system and connected tools determine where data goes. Size the complete application, including working memory for concurrent requests, rather than treating the model's fit claim as a throughput promise.

The model supports configurable reasoning and tool calls, but your application must provide the tools and execute them. It also requires the harmony message format. Both checkpoints configure 131,072 context tokens and accept text rather than images, audio or video. Model card, 120b configuration, 20b configuration.

Muse Spark Open changes the Meta option

Muse Spark Open, from Meta, is the new Meta candidate for multimodal agents. Its 816B total / 42B active parameters and 1M-token context come with text, image, video and audio input and text output. For an agent that must understand a recording, inspect a screenshot and take a tool action, that is a broader input specification than a text-only coding model. Muse Spark Open card.

Meta names Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI as hosted inference partners. It names SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth for inference. The download and catalog routes include Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. A named host is an access option, not a guarantee of identical modalities or production availability. Meta announcement.

Self-hosting remains a substantial commitment. Meta's card lists at least 700GB of VRAM for NVFP4 and 2TB for BF16; its Apple route uses two M5 Ultra Mac Studios with MLX. Those are format-specific configurations, not a claim that an ordinary desktop can run the original weights. Model card and deployment requirements.

The Muse Model License permits commercial uses subject to its conditions. You must retain the required notices, comply with the Usage Policy and display Built with Muse Spark Open in one of the specified places. Section 3 requires notification and a separate written agreement when a licensee or affiliate provides the defined model-as-a-service offering and their aggregate gross revenue exceeds $20 million during any consecutive 12 months, before offering Muse or its derivatives as part of that business. It explicitly excludes internal use, research, other commercial activity that is not model as a service, and features of your own product when that product is not model as a service. This is not a blanket revenue cap on every application built with Muse. Muse Model License.

If your existing deployment already meets the workload, keep it as the baseline rather than migrating on release news alone. For a new Meta evaluation, compare Muse's actual task outcomes and deployment conditions before allocating the hardware.

Who should pick what

The right model follows your binding constraint. Use these eight primary picks as a shortlist, with Small 4 as the Apache-licensed Mistral alternative and gpt-oss-20b for smaller hardware.

For long documents, start with the input format. Text-only repositories can go to GLM; scanned pages need a model that reads images, such as Large 4 or Small 4. Recordings with video favor a Kimi or MiniMax evaluation, while native audio input is a reason to consider Muse. Test whether answers point to the correct passage, page or event instead of choosing solely by the context limit.

For small hardware, start with gpt-oss-20b's documented 16GB memory target. If image input and Apache 2.0 are mandatory, Small 4 is an alternative to evaluate on appropriately sized infrastructure; its 119B total parameters do not make it a normal-laptop model.

Size the deployment before you switch

Choose the model and the serving configuration together. A downloadable checkpoint has no universal per-token price or universal GPU fit. Your precision format, context length, concurrency and connected tools determine whether the application fits the infrastructure you control.

1. 1

   Evaluate a representative workload

   Use representative tasks and record accepted outputs, latency, retries and generated tokens. A hosted trial can narrow the shortlist when your data policy allows it; make the final decision on the checkpoint and serving configuration you will deploy.

2. 2

   Pin the checkpoint and its interface

   Record the full repository ID, revision, license, precision and inference engine. Keep its message format and reasoning controls with that record. Flash-1017's reference encoder, gpt-oss's harmony format and M3.1 Flash's effort settings are part of the integration, not interchangeable API decorations.

3. 3

   Self-host after sizing the complete workload

   Budget for the weights, request memory, concurrency, idle capacity and engineering time. Start with the model's documented format and serving recipe, then measure your workload. Active parameter count describes computation; it is not the amount of storage or memory you can assume you need.

There is no defensible universal savings multiplier here. Self-hosting may earn its cost through control or steady utilization; a hosted API may be cheaper for intermittent work. Compare completed jobs and the operating burden. For the wider coding choice, see the best AI model for coding.

The decision rule

Decide the deployment boundary first. If you can use an API, compare capability, host terms and the bill on your own tasks. If the model must run on infrastructure you control, filter by the exact license and a realistic hardware configuration before comparing benchmarks.

The next move is concrete: take a representative batch from your product, run your shortlisted model and current baseline with the same tools, and inspect the failures. Keep the model that meets your quality threshold within the deployment budget. A release earns a migration by improving that result.

What is the best open-source LLM in 2026?

There is no single winner across these deployment constraints. Evaluate GLM-5.3 and Mistral Large 4 for coding agents, DeepSeek V4.1 Flash-1017 when MIT weights matter, MiniMax M3.1 Flash for long multimodal workflows, and gpt-oss-120b for local text reasoning. Muse Spark Open adds multimodal inputs and broad deployment support, with custom commercial terms.

Is open weight the same as open source?

No. Open weight describes access to trained parameters. The license determines permitted uses, and the Open Source AI Definition also requires code and training-data information. Check the exact release: a family's license can change between versions.

Can I run these models on my own hardware?

Yes, if the checkpoint's license and serving requirements fit your deployment. The released gpt-oss-120b quantization targets one 80GB GPU; the larger models need their own capacity plan. Muse's card describes substantially larger configurations, including its two-Mac-Studio route. Tool connections and logs still need their own data controls.

What is the best open LLM I can run locally on a normal machine?

gpt-oss-20b is the starting point here when you have the documented 16GB of memory and want text reasoning under Apache 2.0. Check the runtime and available memory before choosing it; the larger flagship models in this list are a different deployment class.

Want the full landscape mapped to your business, not just the models but the agents, tools, and stacks that turn them into products? Get the AI tools map for business owners and a short weekly read on what actually shipped and what to do about it.