Open-weight AI models in 2026: a business buyer's guide

Open-weight AI models are ready for business use in 2026. The best of them — Kimi K3, GLM-5.2, Kimi K2.6, DeepSeek V4, Qwen3.5, OpenAI's gpt-oss-120b, Mistral Large 3 — now sit weeks rather than months behind the closed frontier, and two of the strongest ship under MIT or Apache 2.0, licences that permit commercial deployment without asking anyone's permission. The catch is not capability. It is reading the licence, sizing the memory, and matching the model to the job.

What does open-weight actually mean, and why is it not open-source?

An open-weight model is one whose trained parameters — the weights, a file of hundreds of billions of numbers — are published for anyone to download, run and fine-tune on their own hardware. Open-source, used strictly, means more than that: the training code, the data recipe and an OSI-approved licence. Almost nothing at the frontier clears the stricter bar. Nearly everything useful ships as open weights plus a legal document, and the legal document is where deployability gets decided.

The distinction is not pedantry; it is the thing that decides legality. "Open" in a model announcement works like "free" in enterprise software — a marketing word until you have read the terms. Two models can both be described as open while one arrives under MIT, which lets you do essentially anything, and the other arrives under a custom "community" licence carrying usage caps, field-of-use restrictions or territory clauses. Meta's Llama licence, with its user-count carve-outs, was for years the canonical example of open-with-strings — and Meta has now exited open weights altogether, which tells you how little the word "open" ever guaranteed by itself.

So the buyer's translation is short. You are acquiring two artefacts: a very large file and a licence. The file determines what the model can do. The licence determines what you may do with it. Evaluate both; only one of them can get you sued.

Which open-weight models lead the field in 2026?

The open-weight field is better held as durable tiers than as a league table. Benchmark rankings go stale in weeks; the tiers below have been stable for months even as names shuffle inside them, and they map directly onto hardware budgets and business jobs.

As of late July 2026, Kimi K3 — 2.8 trillion parameters, weights released on 27 July under a modified-MIT licence — tops the open-weight rankings outright, while GLM-5.2, the MIT-licensed 753-billion-parameter model that led for most of the summer, remains the strongest open model that fits on a single box. The gap between the best open models and the closed frontier has collapsed from months to weeks.

The release that settled it: Moonshot AI announced Kimi K3 on 17 July, and the weights landed on 27 July exactly as promised — 96 safetensors shards totalling 1.56 TB in native MXFP4, with the licence file beside them. Artificial Analysis scores K3 at 57 on its Intelligence Index, the highest open-weight result it has recorded, six points clear of GLM-5.2 and within touching distance of the best closed models. The licence reads clean for most businesses: a Kimi K2-style modified MIT under which internal use is entirely unrestricted, with conditions only for model-as-a-service resellers above $20 million in revenue and consumer products above 100 million monthly users. By this guide's own rules, K3 was deployable from day one. The catch has moved from paperwork to physics: 1.56 TB of weights outgrows every single machine you can buy.

Beneath that headline, three tiers:

• •The frontier tier: reasoning heavyweights. GLM-5.2; Kimi K2.6, a trillion-parameter model whose INT4 quantisation still occupies about 594 GB; DeepSeek V4 Pro — with Kimi K3 now heading the tier — 2.8 trillion parameters, 1.56 TB in native MXFP4, multi-node territory from day one. Reach for these when the job is genuine multi-step reasoning — agent workflows, long document chains, serious code.
• •The workhorse tier: volume and speed. DeepSeek V4 Flash trades some depth for throughput; Qwen3.5-397B is the big multilingual all-rounder. High-volume operational work — extraction, classification, triage at scale — lives here.
• •The single-box tier: capability on a desk. Qwen3.5-122B, OpenAI's gpt-oss-120b and Mistral Large 3 run on desk-side hardware. Less headroom, dramatically less infrastructure.

The market's shape matters as much as the names. The top of the field is dominated by Chinese labs — Zhipu's GLM, Moonshot's Kimi, DeepSeek, Alibaba's Qwen — with OpenAI back in the category via gpt-oss and Mistral carrying the European flag under Apache 2.0. Meta, which popularised open weights, no longer publishes them.

And the business translation of "weeks behind": for operational workloads — triage, extraction, drafting, agentic back-office flows — that gap is invisible in practice. You are not buying bragging rights. You are buying, permanently and on your own hardware, the capability level the closed labs were selling at the start of the year.

Which licences can you actually deploy under?

Two model licences need no legal review to speak of: MIT (GLM-5.2) and Apache 2.0 (gpt-oss-120b, Mistral Large 3). Both are unrestricted for commercial use — deploy, modify, fine-tune, embed in a product, redistribute — and Apache 2.0 adds an explicit patent grant your counsel will appreciate. If a model you want ships under either, licensing is a solved problem.

Everything else ships under a custom licence, and custom means read it. The licence file sits next to the weights and is rarely longer than a few hundred lines; reading it is the cheapest legal due diligence you will ever do. Four things to look for:

• •Usage restrictions. Field-of-use bans and bolt-on acceptable-use policies — sometimes ones the vendor reserves the right to update after you have built on the model.
• •Attribution and notice terms. Some licences require a "built with" credit in your product or documentation. Trivial to comply with, embarrassing to discover late.
• •Redistribution and derivative terms. Can you ship a fine-tune to a client? Can a managed-service provider host the weights on your behalf? Custom licences answer these questions differently from one another.
• •Jurisdiction carve-outs. At least one major Chinese-origin model licence has excluded UK users outright. Territory clauses are not a hypothetical drafting quirk; check them before a single engineer builds a dependency on the weights.

Two habits keep you honest. Licences are per release — a family that was permissive at one version can tighten at the next, so never assume continuity. And pin the licence text alongside the weights in your artefact store, so the terms you deployed under are terms you can prove you deployed under.

What hardware runs what?

Hardware for open-weight models comes down to memory first and bandwidth second. A model's footprint is roughly its parameter count times its precision — Kimi K2.6 at INT4 weighs about 594 GB before it has served a single token — and the box either holds that file plus working headroom or it does not; memory bandwidth then sets your tokens per second. That mental model produces three practical hardware classes in the UK right now.

• •The 128 GB unified class, from around £3,211. NVIDIA's DGX Spark Founders Edition (£4,899 inc VAT; 128 GB unified memory at 273 GB/s) runs gpt-oss-120b at roughly 35–40 tokens per second, and Lenovo's ThinkStation PGX is the cheapest door into the class at around £3,211. This tier serves the 120B models — gpt-oss-120b, a quantised Qwen3.5-122B — which is genuinely enough for email triage, summarisation and internal assistants.
• •The 96 GB GPU class, from around £11,333 a card. The RTX PRO 6000 Blackwell (96 GB) sits at about £11,333 inc VAT after a 55% GDDR7-shortage price rise. Two of them in one workstation — £36,999.98 from Scan's 3XS line — is currently the cheapest box that serves DeepSeek V4 Flash natively, which makes it the volume-extraction machine.
• •The 748 GB class, at £117,599.99. DGX Station GB300 machines carry 748 GB of coherent memory; the only UK price we have verified is ASUS's ExpertCenter Pro ET900N G3 at £117,599.99 inc VAT. This is where the true frontier fits on a single box — Kimi K2.6's 594 GB, GLM-5.2's 753B mixture-of-experts — under a desk rather than in a datacentre.

One dead end worth naming: the 512 GB Apple Mac Studio, the budget big-memory route many buyers were counting on, was discontinued in March 2026 amid the RAM shortage. Apple's ceiling is now 96 GB at £5,299, which took the cheap path to frontier-class memory off the map. And one new ceiling worth naming: Kimi K3's 1.56 TB checkpoint is the first open weights to outgrow even the 748 GB class — the smallest on-premises footprint is three networked DGX Stations, rack economics in tower clothing. For a model that size, benchmark the hosted route before buying metal.

We keep the full, priced catalogue at /self-hosted/hardware, and the self-hosted landing page plots the fifteen-model landscape on a capacity-versus-bandwidth chart alongside an interactive TCO calculator. Whether owning the box beats paying per token is its own question — we work the numbers properly in self-hosted AI vs API costs.

Are Chinese open-weight models safe for business use?

Chinese open-weight models are safe to run, with diligence — and the diligence belongs somewhere other than where most people first point.

Start with the mechanism, because it settles the scariest version of the question. Model weights are a static artefact: a very large file of numbers that an inference server you control multiplies against your input. The file contains nothing that executes on its own, and it has no network access beyond whatever the surrounding software is given. Download DeepSeek V4 onto your own hardware, unplug the network cable, and it works identically. That is categorically different from sending data to a Chinese-hosted API or consumer app — a live cross-border data transfer that deserves every hard question it gets.

A downloaded model cannot phone home — the China question lives in the licence file and the training data, not in the network traffic.

The genuine considerations are these:

• •Licence and jurisdiction terms. As above: one major Chinese-origin licence has already excluded UK users, so read territory clauses first, not last.
• •Training-data provenance and bias. You cannot audit the corpus. For invoice extraction it rarely matters; for public-facing or politically adjacent content, outputs can reflect the training regime — so eval the model on your own domain before trusting it there.
• •Perception and procurement. Your regulator, your customers and your insurer may hold views that no technical argument will move, and some sectors will simply say no. Decide with eyes open and write the decision down.
• •Supply-chain hygiene. Fetch weights from the official repository, prefer safetensors files (inert data, unlike older pickle formats), verify checksums and pin versions — the same discipline you would apply to any third-party artefact.

The strongest containment available is the deployment model this whole guide assumes: weights running on your own hardware, inside your own network, in your own tenant — air-gapped, if your risk appetite demands it, because nothing about an open-weight model stops working offline. That posture is how Usermode stands these systems up, with zero customer data held on our side, and it is stronger containment than any assurance about anyone else's cloud.

Which model should you run for which job?

Choose an open-weight model by starting from the job, not the leaderboard. The expensive mistake of 2026 is not picking the second-best model; it is buying frontier hardware for a workload a 120B model clears easily — or the reverse, pointing agent workflows at a model that cannot sustain multi-step reasoning. The picker, priced against the hardware classes above:

Three notes on using the table honestly. Pick the smallest model that clears your quality bar on your own documents — evals, not vibes — and spend the savings on memory headroom. Treat the licence column as a prompt, not an answer: "check per release" means exactly that, because families change terms between versions. And remember the model is only half the deployment: a model answers questions, but to do work — read the mailbox, update the ledger, file the document — it needs governed access to your systems, which is the job of the Model Context Protocol. We explain that layer in what is MCP.

The short version: the models are ready, the permissive licences are genuinely permissive, and the hardware starts at about £3,211. If you would like to see an open-weight model stood up on its own hardware and doing governed work inside a real business, book a demo.