Best Open-Weight AI Models in 2026

Two years ago “best open-weight model” meant “the best model you settle for when you cannot pay for a real one.” That framing is dead. In July 2026 the leading open-weight model, Zhipu’s GLM-5.2, sits level with a closed frontier model on an independent economic-task benchmark: Artificial Analysis scores GLM-5.2 at 1524 on GDPval-AA v2, effectively tied with GPT-5.5 at 1514. You can download the weights and run them yourself.

So the question changed. It is no longer whether an open-weight model is good enough. It is which one fits the job, what it actually costs once you account for the tokens an agent burns, and whether the license lets you ship what you build.

The best open-weight AI models in 2026

The best open-weight AI model overall in 2026 is GLM-5.2, which leads open models on the Artificial Analysis Intelligence Index v4.1 with a score of 51. For coding agents, pick Kimi K2.7 Code or Qwen3-Coder; for cheap high-volume work, DeepSeek V4; for on-device use, Liquid LFM2.5. Here is the full field.

Model	Maker (country)	License	API price $/Mtok (in / out)	Best for
GLM-5.2	Zhipu / Z.ai (China)	Open weights	$1.40 / $4.40	Overall capability, long-horizon coding
DeepSeek V4 Pro	DeepSeek (China)	MIT	$0.44 / $0.87	Cheap, long-context agentic work
DeepSeek V4 Flash	DeepSeek (China)	MIT	~$0.14 / ~$0.28	Highest-volume, cost-sensitive work
Kimi K2.7 Code	Moonshot (China)	Open weights	$0.95 / $4.00	Long autonomous coding runs
MiniMax M3	MiniMax (China)	Open weights (attribution required)	$0.30 / $1.20	Multimodal (text, image, video)
Qwen3-Coder 480B	Alibaba (China)	Apache 2.0	Free to self-host	Permissive-license coding, self-hosting
Mistral open MoE	Mistral (France)	Open weights (early access)	Not yet public	Non-China enterprise optionality
Inkling	Thinking Machines (United States)	Apache 2.0	$1.87 / $4.68	Most permissive frontier-scale license
Liquid LFM2.5	Liquid AI (United States)	Open weights	Free to self-host	Edge and on-device agents
Cohere North Mini Code	Cohere (Canada)	Apache 2.0	Free on hosted tiers	Non-China permissive coding

One thing jumps out of that table before any benchmark does. Six of the ten come from Chinese labs. The open-weight frontier is still largely a Chinese story, but the non-China entrants are filling in the gaps the Chinese labs left open: Mistral (European enterprise procurement), Liquid (the edge), Cohere (permissive Apache 2.0 coding), and now Thinking Machines’ Inkling, the first frontier-scale entrant from a US lab. We dug into why the economics push that way in why Chinese AI models are so cheap.

Prices in the table are standard API rates, checked against each provider’s own pricing page as of mid-July 2026 (GLM-5.2’s rate comes from live API resellers while Z.ai’s own card is still rolling out). They are the yardstick, not the whole bill. More on that below.

What “open-weight” actually means (and what it doesn’t)

Open-weight is not the same as open-source, and the gap matters the moment you ship a product.

Open-weight means the trained model parameters are published: you can download them, run the model on your own hardware, fine-tune it, and serve it without sending a single token to the maker’s cloud. What you usually do not get is the training data or the full training code, so you cannot reproduce the model from scratch. That is why most people say “open-weight” rather than “open-source.” A handful qualify as closer to fully open, but treat that as the exception.

The license is where the real decisions live. Three tiers show up across this field:

• Permissive (Apache 2.0, MIT). DeepSeek V4 ships under MIT, Qwen3-Coder under Apache 2.0, Cohere’s North Mini Code under Apache 2.0, and Thinking Machines’ Inkling under Apache 2.0 too, the first frontier-scale (975B) model in this tier. Commercial use, modification, redistribution, no royalties, no strings. This is the tier you want if lawyers are involved.
• Attribution-required. MiniMax M3 is free to use, but put it in a commercial product and you must display “Built with MiniMax M3” somewhere user-facing. Fine for many, a dealbreaker for a white-label vendor.
• Custom or evolving. Several Chinese labs release under bespoke community licenses whose commercial terms have changed release to release. The weights are there; the terms are the thing to read before you build.

The practical rule: pick the model on capability and cost, then read the actual license file before you commit, not the blog post announcing it. An open download is not the same as an open right to sell what you make with it.

Best for coding agents

If the job is a coding agent that runs for hours and finishes real software tasks, two picks stand out for different reasons.

Kimi K2.7 Code is Moonshot’s coding-specialist model, released June 12, 2026 with stronger long-horizon coding and about 30% lower thinking-token use than its predecessor. Its whole design goal is completing a software task start to finish without falling apart across thousands of tool calls. On a general intelligence index it looks middling (42 on Artificial Analysis, below GLM-5.2), but that undersells it: a general composite penalizes a model tuned narrowly for agentic coding. Judge it on whether the agent lands the change, not on a physics-question score. And Moonshot’s successor is imminent: a Kimi K3 launch promotion went live around July 15, though the headline specs circulating are press reports, not a vendor announcement, so wait for the weights before betting a workflow on them. We track what is confirmed versus rumored in the Kimi K3 explainer.

Qwen3-Coder 480B is the pick when the license is the constraint. It is open-weight under Apache 2.0 at 480 billion parameters and matches top closed models on agentic coding tasks, and because it is Apache-licensed you can self-host and ship commercially with no attribution requirement. Note the trap that trips up half the “best open source” listicles: Alibaba’s flagship Qwen3.7 Max is closed-weight and API-only. Max is not open. Qwen3-Coder is. Do not confuse them.

GLM-5.2 is the third option and arguably the strongest raw coder of the three, but its edge is being a generalist that also codes well rather than a coding specialist. If you want one model for the whole workflow, it is the safer bet. For how these stack up against the closed tools most teams actually pay for, see the AI coding agents comparison.

Best for cheap, high-volume, long-context work

When you are pushing millions of tokens a day through an API, the model that wins is rarely the smartest one. It is the one that stays cheap without falling off a cliff on quality. That is DeepSeek V4.

The Pro tier lists at about $0.44 in and $0.87 out per million tokens on DeepSeek’s own price page, with a 1M-token context window and, critically, a cache-hit discount of roughly 99%. Repeated context, the system prompt and the codebase an agent re-reads on every step, becomes almost free. The lighter V4-Flash goes cheaper still, near $0.14 in and $0.28 out. Both ship under MIT, so self-hosting is a clean fallback if you outgrow the API.

DeepSeek was also the model that broke the psychological barrier. Artificial Analysis put V4 Pro and V4 Flash back among the leading open weights at a price that made teams stop treating open weights as a downgrade and start treating them as the default. If your workload is high volume and cost-sensitive, start here and only move up if quality forces you to.

Best open-weight multimodal model

Most open-weight models read text. MiniMax M3 also sees.

MiniMax released M3 on May 31, 2026 as a 229.9-billion-parameter mixture-of-experts model with 9.8 billion active parameters, a 1M-token context window, and native image and video input, with the weights published to Hugging Face days later. It scores 44 on the Artificial Analysis Intelligence Index, level with DeepSeek V4 Pro and ahead of Kimi K2.6, and its own reported 59.0% on SWE-Bench Pro would beat GPT-5.5 (that figure is vendor-reported and not independently reproduced, so weight it accordingly).

At $0.30 in and $1.20 out per million tokens, it is also cheap for what it does. The one string is the license: commercial use requires the “Built with MiniMax M3” credit. If native multimodality matters and that credit does not bother you, nothing else open touches it right now.

Best for edge and on-device

Here is the category the frontier labs mostly skip, and the one worth watching. Not everything runs in a data center.

Liquid AI builds device-native models, and its LFM2.5-230M, released June 25, 2026, runs an agentic tool loop on hardware you already own: 213 tokens per second on a Galaxy S25 Ultra, and it runs on a Raspberry Pi. That is a 230-million-parameter model, roughly a thousandth the size of the giants above, doing real structured work at the edge with no network round-trip and no per-token bill.

The economics are different in kind. There is no API rate because there is no API: the cost is the device you already bought. For a phone app, an offline tool, or anything privacy-sensitive that cannot phone home, an on-device open-weight model is not a compromise, it is the only design that works. The trade is obvious: a 230M model is not writing your backend. But for classification, extraction, and tight agent loops on-device, it is a real answer where the frontier has none.

Best non-China and permissive-license options

If your procurement rules exclude Chinese-hosted models, or you simply want a Western supplier, the open-weight field narrows fast. Three names carry it.

Qwen3-Coder is technically Alibaba, but the Apache-2.0 weights are yours to self-host anywhere, which is what most procurement teams actually care about. Cohere’s North Mini Code is Canadian, a 30-billion-total, 3-billion-active sparse model under Apache 2.0 that runs on a single H100 and is free on hosted tiers. Mistral, the French lab, has confirmed a new open-weight mixture-of-experts model entering partner early access in July 2026, described by CEO Arthur Mensch as “fat but sparse.” The public weights are not out yet, so treat it as confirmed-but-pending rather than shippable today.

For the full picture on Western frontier options, including the closed ones, see our best frontier AI models excluding Chinese labs breakdown.

Inkling: the first frontier-scale open-weight model from a US lab

Everything above this point tops out around 480 billion parameters. Mira Murati’s Thinking Machines changed that on July 15, 2026 with Inkling: a 975-billion-parameter mixture-of-experts model (41 billion active) released under a full Apache 2.0 license, weights and an NVFP4 quantized variant included, on Hugging Face. Nothing else in this non-China conversation is close to that scale, and nothing else in the field carries a cleaner license: Apache 2.0 has no attribution string and no “custom or evolving” fine print to read before you ship.

Scale and license are not the same as capability, and Thinking Machines says so itself: its own launch framing calls Inkling “not the strongest overall model available today, open or closed.” The company’s benchmark table backs that up. Inkling edges DeepSeek V4 Pro on AIME 2026 (97.1 vs 96.7) but trails GLM-5.2 and Kimi K2.6 substantially on Terminal Bench 2.1 (63.8 vs 82.7 and 71.3), and its Artificial Analysis Intelligence Index of 41 sits below GLM-5.2’s 51. That is the exact “pick for the job, not the leaderboard” case this piece keeps making: Inkling is not the model to reach for if you want the single best open-weight coder today, but it is the most permissive, largest license-clean base available if you plan to fine-tune rather than use it out of the box, which is precisely how Thinking Machines built it, as a customizable foundation for its Tinker platform. The full pricing and benchmark breakdown is in the dedicated Inkling launch post.

The real cost: per token vs per finished task

This is the number every “best open-weight” listicle gets wrong, and it is the one that decides your bill.

The sticker price is per token. Your actual cost is per finished task. A modern coding agent does not send one prompt and get one answer. It reads files, plans, calls tools, reads the results, reruns tests, and loops, sometimes thousands of times, before it finishes one change. The token count per task can run into the millions. So a model with a low per-token rate but weak tool-calling can loop more, burn more tokens, and cost more per finished task than a pricier model that gets it right the first time.

Read the chart as a value map, not a leaderboard. Up is more capable; left is cheaper. DeepSeek V4 Pro in the lower-left value corner delivers all-but-top capability at the lowest price, which is why it became so many teams’ default. GLM-5.2 buys the top of the field, and you pay for it. Kimi sits low on this axis only because a general index cannot see what it is good at.

To run your own numbers instead of trusting a headline rate, we built two tools for exactly this: the AI coding cost calculator models cost per task from published per-token prices and your own assumptions, and the coding agent cost-per-task tracker keeps modeled per-task figures updated as prices move. If you are weighing self-hosting instead of the API, the self-hosted LLM cost per token breakdown covers the compute math, and how much RAM you need to run a local LLM covers the hardware reality before you buy a GPU. To rank the current field by value per dollar directly, the AI model value leaderboard does the sorting for you.

The sticker price is per token. Your bill is per finished task. Those are not the same number, and the gap is where teams overpay.

Meta went the other way

One release makes the open-weight surge easy to misread as inevitable. It is not.

While the Chinese labs doubled down on open weights through the first half of 2026, Meta shipped Muse Spark 1.1 on July 9 as its first paid model, served through a metered API rather than as free weights. The company that made open weights mainstream with Llama pivoted toward a paid, closed product just as its former imitators made openness their whole strategy. Openness is a business decision, not a law of physics, and it can reverse. We tracked the broader flip in how open-source LLMs overtook proprietary models in 2026.

The takeaway for anyone building on open weights: the models available to download today are a snapshot, not a guarantee. Pick for the job in front of you, keep your integration model-agnostic where you can, and do not assume the door that is open now stays open forever.

Frequently asked questions

What is the best open-weight AI model in 2026?

GLM-5.2 from Zhipu is the best overall open-weight model as of July 2026. It leads open models on the Artificial Analysis Intelligence Index v4.1 with a score of 51 and is competitive with closed frontier models on long-horizon coding. For coding agents specifically, Kimi K2.7 Code and Qwen3-Coder are stronger picks, and DeepSeek V4 is the best value for high-volume work.

Is open-weight the same as open source?

No. Open-weight means the trained model parameters are published so you can download, run, and fine-tune the model yourself. Open source usually implies the training data and code are also released so the model can be reproduced from scratch. Most models people call open source in 2026 are actually open-weight: the weights are free, but the training data and full recipe are not.

Can I use open-weight models commercially?

It depends on the license. DeepSeek V4 (MIT) and Qwen3-Coder and Cohere North Mini Code (Apache 2.0) allow commercial use with no strings. MiniMax M3 allows commercial use but requires you to display a "Built with MiniMax M3" credit. Some Chinese labs use custom community licenses with terms that change between releases, so read the actual license file before you ship.

Are Chinese open-weight models safe to use?

Running open weights on your own hardware sends no data to the model maker, which removes the main privacy concern with hosted APIs. The remaining questions are procurement and compliance policy, not technical: some organizations restrict Chinese-origin models regardless of where they run. If that applies to you, Qwen3-Coder (Apache 2.0, self-hostable), Cohere North Mini Code (Canada), Mistral (France), and Inkling (Thinking Machines, United States, Apache 2.0, released July 15, 2026) are the main non-China alternatives, with Inkling the first of the group at frontier scale.

Which open-weight model is best for running locally?

For a phone or edge device, Liquid LFM2.5 runs an agent loop on hardware you already own. For a workstation or single-GPU server, Cohere North Mini Code runs on a single H100 and smaller Qwen3 checkpoints run on consumer GPUs. For a top-tier model on your own cluster, DeepSeek V4 and Qwen3-Coder ship under permissive licenses that make self-hosting clean.

How much does it cost to run an open-weight model?

Through a hosted API, DeepSeek V4 Pro runs about $0.87 per million output tokens and MiniMax M3 about $1.20, versus $4.40 for GLM-5.2. But the API rate per token is not your real cost. A coding agent can use millions of tokens per finished task, so caching discounts and tool-call reliability change the per-task bill more than the sticker rate. Model the cost per finished task, not per token.

Sources

• Artificial Analysis (2026). GLM-5.2 is the new leading open weights model on the Artificial Analysis Intelligence Index (independent benchmark). artificialanalysis.ai
• Artificial Analysis (2026). Comparison of Open Source AI Models (Intelligence Index v4.1 leaderboard). artificialanalysis.ai
• Artificial Analysis (2026). DeepSeek is back among the leading open weights models with V4 Pro and V4 Flash. artificialanalysis.ai
• Artificial Analysis (2026). Kimi K2.7 Code: Intelligence, Performance and Price Analysis. artificialanalysis.ai
• DeepSeek (2026). Pricing (official API price page; V4 Pro and V4 Flash, MIT license). api-docs.deepseek.com
• DeepSeek-AI (2026). DeepSeek-V4-Pro (model card, MIT license). huggingface.co
• MiniMax (2026). MiniMax M3: Frontier Coding, 1M Context, Native Multimodality (maker blog, released May 31 2026). minimax.io
• TechTimes (2026). MiniMax M3 Open-Weight Coding Model: Frontier Claims, Unverified Benchmarks (secondary; license attribution requirement, vendor-reported SWE-Bench Pro). techtimes.com
• Liquid AI (2026). LFM2.5-230M: Built to Run Anywhere (maker blog, released June 25 2026). liquid.ai
• Cohere (2026). North Mini Code (maker blog; 30B/3B MoE, Apache 2.0). cohere.com
• TechTimes (2026). Mistral AI Targets Frontier Gap With Open-Weight Model Entering July Early Access (secondary; CEO-confirmed). techtimes.com
• Meta (2026). Introducing Muse Spark (maker blog; Meta’s first paid model, July 9 2026). ai.meta.com
• Digital Applied (2026). Qwen Models Complete Guide (secondary; Qwen3-Coder Apache 2.0 vs Qwen Max closed). digitalapplied.com
• Thinking Machines Lab (2026, July 15). Inkling: Our open-weights model (maker blog; Apache 2.0, 975B/41B MoE). thinkingmachines.ai