Five prominent open-weight model families share a label, but not the same route to use, price evidence or licence certainty. Shattered.io’s comparison gives developers practical checks before they commit a model to an API, a cloud catalogue or their own hardware.
Watch Desk analysis
What happened
The comparison covers Kimi K3.1 Pro, DeepSeek V4.1 Flash, Mistral Large 4, GLM-5.3 and Gemma 4. It distinguishes models with downloadable weights from hosted API offerings, and separates facts confirmed by official pages from claims backed only by third-party trackers or single mentions.
That distinction changes the practical picture. Shattered.io says DeepSeek’s official pages confirm its API model and pricing, but do not establish a downloadable weight release or a model-card licence. It says Mistral names its Mistral Ecosystem Open-Weight Licence, while some other models need a closer check of the exact model card before deployment. The article also notes that Z.ai routes requests using older GLM model identifiers to newer models, so existing integrations may be using a different model than their names suggest.
Why it matters
“Open” is not a deployment plan. A team may be choosing between a downloadable model it must host, a cloud catalogue and a vendor API, each with different costs, control and licence terms. Treating those options as interchangeable is a tidy way to acquire an unexpected bill or a compliance question.
The comparison also makes a useful case for checking the exact model identifier and licence in the repository or documentation you plan to use. Names, prices and routes can change; a confident-looking table is no substitute for checking what your own integration will actually call.
Our read
This is a useful buyer’s guide precisely because it does not pretend every blank is a fact. Shattered.io labels what it says is confirmed and what remains unconfirmed, though its conclusions are still the publication’s research rather than a universal audit. Before building around a model, check its current model card, licence and API identifier. Glamorous work? No. Less glamorous than a surprise migration, certainly.
What to watch
- Whether the model makers publish clearer weight-release and licence details.
- Whether API prices and model identifiers change after the comparison’s 13 November 2026 snapshot.
- Whether teams report practical differences between hosted access and self-hosting these models.
Discussion spark: When choosing an AI model, should teams prioritise the freedom to run downloadable weights, or the simplicity of a hosted API even when its terms and costs can change?
Sources and evidence
- Open Source AI Models: 3 Run Paths, 48% Coding Index [2026] – shattered.io (13 November 2026, 07:59 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.