Open weight models: what they are and how to use them

What are open weight models

Open weight models are AI models whose final trained parameters, the weights and biases, are published so anyone can download, run, fine-tune, and redistribute them under the terms of their license. You get the finished model, the lab keeps the code, data, and methods that produced it, and the license decides what you may do with what you downloaded.

What a weight is, in plain terms

A weight is a learned number that sets how strongly one input contributes to the network's calculation, and biases are additional learned values added on top. A model described as having seven billion parameters holds roughly seven billion of these numbers. That is why size matters so much in practice: a trillion-parameter model is a trillion values that must sit in memory before it can answer anything.

What you get and what stays private

A typical release gives you enough to run and adapt the model, and nothing that would let you rebuild it.

• Shipped: the weights, a model card, and often a tokenizer and inference config
• Withheld: the training code, the training data, intermediate checkpoints, and the full methodology

Because the training side is withheld, you cannot reproduce the model or audit where a bias entered it. You can still test and fine-tune it, which covers what most production teams need.

Are open weight models free?

The download is usually free, but running the model costs money, either for GPUs you operate or for a hosted endpoint that charges per token.

Short answer: The weights cost nothing to download, but compute does. The license, such as MIT or Apache 2.0, sets the commercial terms. Even OpenAI, whose flagship GPT-5.5 stays closed, released a downloadable open weight model, gpt-oss, in 2025.

If you would rather call a hosted open weight model than download around 1.5TB of weights, the Telnyx inference models page lists each one with its parameters and context length.

Run open weight models without owning GPUs

Call GLM, Kimi, DeepSeek, and MiniMax models through one OpenAI-compatible request and pay per token.

Try Telnyx Inference

Open weight vs open source vs proprietary models

The three categories differ in what you can see, what you can move, and who sets your price.

Open weight vs open source models

The Open Source Initiative published its definition of open source AI in late 2024. A qualifying model must ship its training code, information about its training data, and its full methodology, and it must grant four freedoms: Use, Study, Modify, and Share. The OSI open weights definition sets out where weight-only releases fall short.

Open weight compared with open source AI, component by component

Almost no popular model clears that bar. Llama, Qwen, DeepSeek, and GLM release weights only, while OLMo and BLOOM, both from research projects, are the open source examples. For agents, the gap limits what you can prove to an auditor more than what you can build. Our roundup of the best open source LLMs shows how vendors use the term, and open source LLMs explained covers the basics.

Open weight vs closed weight models

You reach a closed weight model only through its vendor's API. Staying closed costs you four things:

• Lock-in to one vendor's roadmap and uptime
• No way to inspect the model
• No option to run it on your hardware or in your region
• A price the vendor sets and can change

Closed frontier models still lead on some tasks, which is a fair reason to keep them for part of your traffic. Open weights carry their own trade-off: once published, they cannot be withdrawn, and anyone can fine-tune the safeguards out. A 2024 NTIA report weighed that risk and recommended continued monitoring over immediate restrictions.

What each open weight license lets you do

The license decides what you may do commercially, so read it before the model page. Four license types cover the current leaders.

Open weight licenses before commercial use

Both MIT and Apache 2.0 allow commercial use. Apache 2.0 adds an express patent license, which ends if you sue a contributor over patents.

Modified MIT adds terms to a permissive base, and custom community licenses can restrict uses or scale, so read the LICENSE file in the repository rather than the label. Professors Peter Henderson and Mark Lemley argue these terms may be largely unenforceable, because weights are machine-generated and functional and may carry no intellectual property rights. Follow the license anyway.

Regulated workloads: In healthcare, financial underwriting, or public administration, final weights alone may not satisfy fairness, privacy, or explainability requirements, so plan extra documentation. For support, sales, and scheduling agents, open weights are usually enough.

Best open weight models available now

As of October 2026, Moonshot AI's Kimi K3 is the largest model on the Telnyx models page at 2.8 trillion parameters, well past DeepSeek V4 at 1.6 trillion, which older roundups still call the largest. Telnyx hosts it, as covered in Kimi K3 on Telnyx, but the picker below groups models by job, since size says little about fit.

Open weight model picker, October 2026

Best for agents and long-horizon coding

GLM-5.2 from Z.ai tops the independent Artificial Analysis Intelligence Index for open weight models and ships under MIT with a 1 million token context. Against a closed frontier model it gives up input types and footprint: it reads text only, and the full weights run to around 1.5TB. Telnyx also hosts GLM-5.3 and GLM-5.3 Flash, which makes GLM the first family to test for retail and e-commerce agents.

Kimi K2.7 Code, released by Moonshot AI in June 2026, spends 30% fewer thinking tokens than K2.6 and accepts image input. Moonshot's own tests show it beating Opus 4.8 on MCP tool use.

Who ran the benchmark: The Artificial Analysis index is independent. Most other headline scores for GLM, Kimi, and MiniMax come from their publishers, so test them on your own prompts.

Best for reasoning and high-volume tasks

DeepSeek V4 Pro has 1.6 trillion parameters and puts reasoning first, which burns output tokens. It can also answer when it does not know or drop requirements from prompts with many constraints.

DeepSeek V4 Flash runs at around 112 tokens per second with three reasoning modes, from thinking off to max thinking. It suits tagging tickets and pulling fields out of emails, and Telnyx Inference charges $0.13 per 1M input tokens and $0.26 per 1M output tokens for it at default rates.

Best for multilingual and multimodal work

Qwen3.5 397B from Alibaba is a mixture-of-experts model with 397B total parameters and 17B active per token. It covers 201 languages and dialects and accepts text and images, though coding trails the leaders slightly.

MiniMax M3, released in June 2026, is the one to test if your agent must read video. It accepts text, image, and video, and its sparse attention processes million-token prompts over 9x faster than its predecessor. Kimi K2.6 adds image and video input and swarms of up to 300 sub-agents, by Moonshot's figures.

Freshness: Parameter counts and largest-model claims change monthly. These figures were checked in October 2026, and each model card on Hugging Face is the source for its specs.

Open weight models outside China

If procurement rules out Chinese labs, US and European publishers still give you options. Thinking Machines Lab's Inkling, at 975B parameters, is the largest, and Meta's Llama has the longest tooling history.

gpt-oss (OpenAI)

For teams on OpenAI's API that want a downloadable model from the same publisher.

Mistral Large 3

675B parameters under Apache 2.0, for teams that prefer a European publisher.

Nvidia Nemotron 3 Ultra

A 550B model for teams that want the GPU vendor's own release.

Gemma, Phi, and Granite

Smaller families from Google DeepMind, Microsoft, and IBM for narrow, high-volume steps.

Small open weight models for a laptop or one GPU

Not every agent step needs a trillion parameters. Gemma 4 runs from 2.3B to 30.7B parameters, Microsoft's Phi-4 is a 14B model under MIT, and IBM's Granite family rounds out the group. These suit classification, field extraction, and on-device assistants.

On Telnyx: Gemma 4 26B A4B, a 25.8B model with 256K context, costs $0.13 per 1M input tokens and $0.40 per 1M output tokens at default rates, which makes it a cheap router or classifier in front of a larger model.

How to choose an open weight model

Four questions decide the choice: what you can run, what each task costs, how the model does on your prompts, and what its model card says. Our guide to open weight model selection covers more trade-offs.

How big a model can you actually run?

Total parameters set the memory you need to hold a model, and active parameters set the compute it spends per token. Full GLM-5.2 weights are around 1.5TB. MiniMax M3, at around 428B parameters, needs multi-GPU hardware. Qwen3.5 runs 397B total but only 17B active per token, so it computes like a small model while needing the memory of a large one. Kimi K2.6 ships natively quantized at 1 trillion total and 32B active. Quantization stores each weight in fewer bits, cutting memory at some cost to quality. A longer context window adds KV-cache memory on top, so a model that fits your GPUs at 8k context can run out of room at 262k.

Measure cost per task, not price per token

DeepSeek V4 Pro's reasoning burns output tokens, so its cost per task can exceed what its per-token price suggests, and DeepSeek V4 Flash is wordy enough to push the same way. Our piece on efficient model selection shows how to weigh that cost against quality.

What to log: For every test task, record input tokens, output tokens including reasoning, and wall-clock time. Multiply the tokens by each provider's published rates, and you have a cost per task to set against your current bill.

Build a small eval set from your own prompts

Vendor benchmarks measure the vendor's tasks, and a two-week test on yours tells you whether you can move traffic.

1. Pick 30 to 50 real prompts from your production logs.
2. Include tool-calling prompts and prompts with long constraint lists, because models drop requirements when a prompt stacks many of them.
3. Score outputs blind against your current closed model.
4. Compare latency at your real concurrency, since a model that answers fast alone can slow down under load.

Check the model card before you download

On Hugging Face, a model card is the repository's README.md with metadata. It states intended uses, limitations, training data notes, and evaluation results, including who ran them. Download weights only from the publisher's official repository.

Hosted and downloaded can differ: Qwen3.5 offers 1M context only in Alibaba's hosted version, and Kimi K2.6 video input works only through Moonshot's API. Test the exact artifact you will deploy.

How to run inference on open weight models

You can run open weight models on your own GPUs or through a managed endpoint, or keep a closed API where no open model matches. Volume, data rules, and spare engineers decide which, and our overview of model deployment options covers the wider set of patterns.

Should you self-host or call a managed endpoint?

• No GPUs, want a hosted model today: call Telnyx Inference and pay per token. You skip pulling around 1.5TB of weights and never buy hardware.
• Own GPUs, fine-tuning on private data: download a seven billion parameter model and self-host it. It fits in memory far more easily than a trillion-parameter one.
• Moving agent traffic off a closed API: trial Kimi or DeepSeek on Telnyx Inference with one OpenAI-compatible request. Compare cost per task on your prompts before migrating.
• Shipping a paid product on the weights: pick a model under MIT or Apache 2.0 before choosing a host. The license, not the free download, sets your commercial rights.
• Building a phone voice agent: run inference on Telnyx next to speech to text and text to speech. That full chain decides the delay a caller hears.

Self-host, managed endpoint or closed API

Self-hosting means you download the weights, provision GPUs, run a model serving stack such as vLLM, TGI, or llama.cpp, and own uptime, scaling, and patching. Anything near 428B parameters needs multi-GPU hardware, and full GLM-5.2 takes around 1.5TB of storage before you serve a token.

Three ways to run open weight models

A managed endpoint hosts the weights and bills per token. Telnyx Inference runs open weight models on Telnyx GPUs in five regions across four continents, with no minimums or commitments, and lists per-model rates on its inference pricing page. At low volume the managed endpoint wins because idle GPUs cost money. At very high, steady volume, self-hosting can win if you have the engineers.

Keep data in your own region

With self-hosting, prompts never leave your environment. With a managed endpoint, you need to know where the provider runs inference and what it stores.

Data residency on Telnyx: Add a region of USA, EU, AUS, or UAE to a request and set the mode to strict, and Telnyx fails the request rather than serving it elsewhere. The chat completions endpoint stores no request or response data. The inference regions page lists the routing rules.

Call an open weight model through an OpenAI-compatible request

Most open weight endpoints accept the OpenAI chat-completions request shape, because existing SDKs and agent frameworks already send it. Switching providers means changing three fields: the base URL, the API key, and the model name. If you have a developer, this is the documented Telnyx request, using Kimi K2.6 from the Telnyx inference models list with an EU region preference:

Coming from Claude: Telnyx also exposes an Anthropic-compatible Messages endpoint at /v2/ai/anthropic/v1/messages. If your agents use the Anthropic SDK, you change the base URL and pass your Telnyx API key as a Bearer token, and the request body stays the same.

Open weight models for voice AI

A voice call runs through a chain: the caller's audio, speech to text, the LLM, text to speech, and network delivery back to the caller. Once the LLM is an open weight model of comparable quality, the latency the caller hears comes from how many vendors and network hops sit in that chain.

Why the model is the easy part

Every vendor boundary adds a network hop, and every hop adds dead air. Telnyx keeps the chain co-located with telephony: it hosts seven TTS engines on its own GPUs and routes to seven more through the same API, so speech to text, inference, and text to speech stay on one infrastructure.

On Telnyx Voice AI, Kimi K2.6 and GLM-5.2 are verified for voice calls, and reasoning is always disabled on calls so thinking tokens never delay a reply. You can also connect any OpenAI-compatible LLM, build with the Telnyx LiveKit model docs, compare TTS model architectures, or see the contact center voice AI setup.

Where open weight models still lag

Open weight models are text-first. GLM-5.2 and both DeepSeek V4 models read text only. Among the models above, image input exists on Qwen3.5, MiniMax M3, Kimi K2.6, and Kimi K2.7 Code, and video input only on MiniMax M3 and Kimi K2.6.

Plan for the gap: Kimi K2.6 video is experimental and available only through Moonshot's API. If your voice agent must read images or video, route those turns to a multimodal model and keep the rest on open weights. Text-only voice agents can move now.

The gap is narrowing. The Telnyx models page already lists DeepSeek V4.1 Flash and Qwen3.8 27B for multimodal work, so recheck it before you commit to a long-term hybrid setup.

Open weight models FAQ

Can we move our agent traffic from a closed model API to open weight models without the output quality dropping?

Often yes for text workloads, and a two-week test on your own prompts settles it. Score two or three candidates blind against your current model, watch tool calling and constraint following, and move traffic one task at a time.

How far behind are open weight models on image and video input?

Furthest behind on video. Several models take images, but video works natively only on MiniMax M3 and experimentally on Kimi K2.6 through Moonshot's API. Text-only agents can move now, and multimodal agents need a routing plan.

Are open weight models free to use commercially?

Usually, within the license terms. MIT and Apache 2.0 allow commercial use with notice requirements, while modified MIT and custom licenses add conditions. The download is free, and compute is what you pay for.

What is the difference between open weight and open source models?

An open weight model publishes its trained weights and keeps its training code and data private. An open source model, under the Open Source Initiative's definition, also ships the training code, training data information, and methodology.

Can an open weight model be withdrawn once the weights are released?

In practice, no. Once weights are widely downloaded, the publisher cannot pull them back, and copies can be modified to remove safeguards. The NTIA reviewed this risk in 2024 and recommended monitoring rather than immediate restrictions.

Are open weight licenses legally enforceable?

That is unsettled. Henderson and Lemley argue the terms assume rights that may not exist in machine-generated weights, and a 2020 Intellectual Property Owners Association whitepaper proposed a new right in trained models. Follow the license regardless.

Build on open weight models with Telnyx

You now have a path from a closed model bill to open weights. Pick two or three candidates from the best-for groups, run the two-week eval on your own prompts, and choose self-hosting or a managed endpoint by volume and data residency.

Teams that go further fine-tune the weights on their own domain data and keep the resulting model.

Start with the Telnyx inference models page to see what runs today, then send the request from the inference section with one candidate.