What open weights let you do, and what they don't

An open-weight model is one whose trained weights anyone can download, run, and fine-tune.

As of October 29, 2026, the strongest one on our overall ranking is Moonshot's Kimi K3. It ranks 17th with 69.3, against 86.3 for Claude Opus 5.5, the top supported closed model.

So open weights rarely buy you the best scores. They buy control over where the model runs, which version you keep, and what you change.

Two things limit that control. Most leading open models need datacenter hardware, and each license can attach conditions to commercial use.

Open weights means the trained numbers are public

Every trained language model is a large file of numbers, the weights, plus the code that runs them. With a closed model, that file stays on the vendor's servers, and you reach it through an app or an API. With an open-weight model, the file sits on a download page, usually Hugging Face, under a license.

That's narrower than "open source". OSI, the Open Source Initiative, published its Open Source AI Definition 1.0 on October 28, 2024, and it asks for three things: the complete code used to train and run the model, the weights including checkpoints, and enough information about the training data for a skilled person to build a substantially equivalent system. OSI's own page on open weights says they differ significantly from open source AI, because the training code and data transparency are usually missing. Stanford HAI's definition is the plain one most people use: the model's core components are released so anyone can download them.

Table 1 Download CSV

What is released	Closed model	Open weights	Open source AI (OSI 1.0)
Access through an app or API	Yes	Usually, through hosts	Usually
Trained weights	No	Yes	Yes
Inference code	No	Usually	Yes
Training code	No	Rarely	Yes
Training data, or a detailed description of it	No	Rarely	A detailed description
License	Terms of service	Varies: Apache 2.0 to custom	Must allow use, study, modification, and sharing

European law sits between the two. Article 53(2) of the EU AI Act relaxes some documentation duties for general-purpose models released under a free and open-source license with public weights. That exemption doesn't cover models with systemic risk, and even exempt models still owe a training-data summary and a copyright policy.

We use the plain definition in our catalog. Once its weights are downloadable, a model counts as open weight, whatever the license says. We then record the license separately, because that's where the differences that matter for a business show up.

The best open-weight model ranks 17th on our board

On October 29, 2026, no open-weight model reached the overall top 16. Kimi K3 is the highest at 17th. It trails the closed leaders by about 17 points on the supported rows and by 19 points against the top estimated row.

Table 2 Download CSV

Ranking, October 29, 2026	Top closed model	Score	Top open-weight model	Rank	Score
Overall	Gemini 4 Carbon (estimated)	88.4	Kimi K3	17	69.3
Coding	Claude Opus 5.5	82.9	DeepSeek V4.1 Flash 1017	14	62.9
Agentic	Claude Opus 5.5	90.1	GLM-5.3	11	69.9
Knowledge	Claude Opus 5.5	89.5	Kimi K3	24	67.2

On agentic work, an open model gets closest by rank. GLM-5.3 from Z.AI sits 11th, ahead of Claude Sonnet 5.5 at 12th. On coding, DeepSeek V4.1 Flash 1017, released October 17 under MIT, is the leading open entry at 14th.

We're waiting on one model in particular. Kimi K3.1 Pro, which Moonshot released on October 23, already ranks 13th overall at 74.6 (estimated) and 9th on agentic work. Moonshot says its full weights will be public by November 4, 2026. Until the files are on Hugging Face, we list K3.1 Pro as pending rather than open. If the weights ship as promised, it becomes the best open-weight model on the board, four places above K3, and we'll update this page when they land.

Depth is better than the top suggests. Thirty-four of the top 100 overall models are open weight, and our catalog holds 528 open-weight entries out of 1,009. In models released in 2026 so far, we track 223 open-weight releases against 171 proprietary ones. One clarification, though: those counts describe what we track, which leans toward models that publish benchmark results. They don't measure the whole market.

While building that table we found a labeling error of our own. Our catalog still lists Kimi K3's weights as pending, a status set when Moonshot announced them. Moonshot's model card on Hugging Face shows the full weights under the Kimi K3 License. We corrected the label on October 29, so K3 now counts as open weight in the catalog and in the tables here.

Scores here are point estimates with wide intervals. Two neighbors two points apart aren't in an established order, so treat the table as tiers, and use the open-source ranking for the full list.

The leading open models need a datacenter, not a laptop

Downloading a model and running it are different problems. Holding the weights alone takes roughly the parameter count times the bytes stored per parameter: 2 bytes in BF16, 1 byte in 8-bit formats such as FP8, and about half a byte in 4-bit formats. Hugging Face's guidance adds up to 20% for inference overhead, plus a cache that grows with the length of the conversation.

Table 3 Download CSV

Model	License	Total parameters	Precision	Weights alone (our arithmetic)	What it takes
gpt-oss-20b	Apache 2.0	21B	MXFP4	about 11 GB	Runs within 16 GB of memory (OpenAI)
Gemma 4 31B	Apache 2.0	31B	BF16 / 4-bit	about 62 GB / 16 GB	One desktop GPU at 4-bit (our estimate)
gpt-oss-120b	Apache 2.0	117B	MXFP4	about 59 GB	One 80 GB GPU (OpenAI)
GLM-5.3	GLM-5.3 License	753B	FP8	about 750 GB	A multi-GPU server (our estimate)
Qwen3.8-2.4T-A95B	Qwen3.8-Max License	2.4T	4-bit	about 1.2 TB	A multi-GPU cluster (our estimate)
Kimi K3	Kimi K3 License	2.8T	MXFP4	1.56 TB (vLLM recipe)	At least 8x GB300; Moonshot suggests 64 or more accelerators

Most of these are mixture-of-experts models, which adds a trap. Kimi K3 activates 104 billion parameters per token, which sounds like a mid-sized model, but any token can be routed to any of its 896 experts. All 2.8 trillion parameters must sit in memory. Active parameters set the compute per token, and total parameters set the hardware bill. gpt-oss-120b shows the same thing at a smaller scale: 5.1 billion active parameters, and still a full 80 GB GPU to hold the rest.

Quantization lowers the floor. Stored at 4 bits, a model needs about a quarter of the memory of BF16, and tools such as llama.cpp, Ollama, and LM Studio run those builds on ordinary machines. You pay for it in accuracy, which can drift from the published scores, which were usually measured at the vendor's own precision.

Our VRAM calculator does this arithmetic for a specific model and context length, and the local LLM shortlist covers the models that fit on one machine.

Most people use open weights through a host

If you can't run a 2.8-trillion-parameter model yourself, someone else can. Inference hosts load open checkpoints on their own GPUs and sell access per token, often with an API that copies OpenAI's format. That's the practical answer to "open but too big", and it creates something closed models don't have: price competition for the same weights.

We record the per-token rates of ten hosts. On October 20, 2026, the spread for the same model looked like this:

Table 4 Download CSV

Model	Host listings	Output rate per million tokens, lowest to highest
DeepSeek V4.1 Flash	14	$0.12 to $1.20
gpt-oss-120b	10	$0.17 to $0.75
GLM-5.3	7	$1.40 to $4.40
Kimi K3	8	$11.94 to $15.00

The cheapest listing is often a router or a promotional route with lower rate limits, so the low end isn't always a production price. Hosts also differ in context length, speed, and the precision they serve, and a host running a 4-bit build may not match the vendor's numbers. Before you switch hosts to save money, rerun your own test set on the new route.

Every rate we track is on the inference hosts page. Our self-host calculator compares renting a host with buying or renting GPUs, and self-hosting only wins when your volume keeps the hardware busy.

The license decides what you can actually do

"Open" says nothing about commercial terms. Of the 528 open-weight models in our catalog, we record a license for 152, and 107 of those use Apache 2.0 or MIT, which allow commercial use with no caps beyond keeping the notice. Several large flagships, including GLM-5.3, Kimi K3, and Qwen's biggest model, use custom licenses that read like MIT until a size or business-model clause kicks in.

We read the license files of the families people ask about most. These are the terms as of October 29, 2026:

Table 5 Download CSV

Model family	License	Commercial use	The condition that matters
gpt-oss-120b and gpt-oss-20b	Apache 2.0	Yes	Comply with applicable law
DeepSeek V4.x	MIT	Yes	Keep the copyright notice
Gemma 4	Apache 2.0	Yes	None beyond Apache
Gemma 1 to 3n	Gemma Terms of Use	Yes, with conditions	Google may restrict use remotely, and the restrictions flow down to models distilled from Gemma
Llama 3.3 and Llama 4	Llama Community License	Yes, with conditions	Companies with over 700 million monthly users need Meta's permission. Derived models must start their name with "Llama", and Llama 4's multimodal rights exclude EU-based companies
GLM-5.3	GLM-5.3 License	Yes, with one condition	API resellers with over $10 billion in yearly revenue need Z.AI's security review
Kimi K3	Kimi K3 License	Yes, with conditions	API resellers over $20 million in yearly revenue need a separate agreement. Large products must display "Kimi K3"
Qwen3.8-Flash-Next	Qwen Community License 1.0	Yes, with conditions	Any API reseller or AI coding or office assistant business needs a separate license
Mistral Medium 3.5	Modified MIT	Only below $20 million monthly revenue	Larger companies need a commercial license

Two patterns stand out. Most restrictions target businesses that resell the model as an API, so a company running a model for its own staff is usually untouched. And the same vendor can use different licenses for different sizes: Qwen3.8-27B is Apache 2.0, while Qwen's flagship carries a custom license.

This is a summary of terms, not legal advice. Licenses change between versions, so read the LICENSE file in the exact repository you download.

Five checks before you build on an open model

Most problems with open models show up after the download, when a license clause or a memory limit surprises someone. These checks take an afternoon and catch the usual ones.

1. Read the LICENSE file in the exact repository. Look for three things: a revenue or user threshold, a rule about reselling access as an API, and a naming or display requirement. Model families change terms between sizes and versions, so the license of a sibling model tells you nothing.
2. Multiply total parameters by bytes per parameter. Use total, not active, parameters for mixture-of-experts models. If the result is larger than your GPU memory with 20% to spare, plan on a host or a smaller build.
3. Match the precision to the published scores. Vendors usually report benchmarks at their own precision, often BF16, FP8, or a native 4-bit format. A community 4-bit conversion of a BF16 model can score differently, so the leaderboard number may not describe the file you run.
4. Compare hosts on more than price. Check the context length, the precision each host serves, its rate limits, and whether the route is a router reselling another provider. Two listings for the same model name can be different builds.
5. Pin the revision and keep your own test set. Record the commit hash you tested, and rerun a few dozen of your real tasks whenever you change the version, the precision, or the host. Our custom benchmark guide shows how to build that test set.

Those five checks also answer the question people usually start with, which is whether an open model is good enough. The ranking tells you how a model does on public tests at the vendor's settings. Your test set tells you how the build you can actually afford does on your work.

What open weights give you that an API can't

Given the benchmark gap, the case for open weights rests on control. Four reasons come up most often.

You can fine-tune and keep the result. LoRA trains a small set of adapter weights on top of a frozen model, and the adapter file is yours. OpenAI says gpt-oss-20b can be fine-tuned on consumer hardware and gpt-oss-120b on one H100 node. Some closed APIs offer hosted fine-tuning, but the tuned model stays on their servers.

Your data can stay on your hardware. A model running on your own machine or private cloud sends nothing to a vendor. Tools such as LM Studio work fully offline once the files are downloaded. That matters for regulated data, source code, and anything under a client confidentiality agreement.

You can pin a version forever. A Hugging Face download can be locked to an exact commit, so the model you tested is the model you run next year. Closed models retire on the vendor's schedule. OpenAI's deprecation page lists gpt-5-2025-08-07 for shutdown on December 11, 2026, and we track the dates for every provider in our deprecations tracker.

You can switch who serves it. The same checkpoint runs on many hosts, so a price change or an outage at one host isn't the end of your setup.

None of that makes an open model the right default. If your task needs the strongest model available and your data can leave the building, a closed frontier model still leads every ranking above. If you need control, privacy, or a fixed version, check the license first and the hardware second. Then measure whether the score gap shows up on your actual work.

We're watching one number next: the rank of the best open-weight model. It was 17th today. If Kimi K3.1 Pro's weights arrive by November 4, it moves to 13th, and whichever release reaches the top 10 will show up first on the open-source ranking.

---