What is the difference between open weight and fully open AI models?

An open-weight release publishes the finished model weights plus a license — enough to run, fine-tune and self-host. A fully open release adds the training lifecycle: intermediate checkpoints, the training data or the recipes used to construct it, the training code, the data-mixture composition and training logs. Open weights make a model runnable; the full record makes a training run reproducible.

Which 2026 models are actually fully open?

Very few. IFM's K2 Horizon fleet, released September 3, 2026 under Apache 2.0, is the most prominent attempt — but even there only the 3.7B and 7B ship the complete artifact set today, while the 375B and 36B cards say checkpoints, data and code will follow. DeepSeek V4 (MIT), GLM 5.3, Qwen3.8 and Kimi K3 are open-weight releases, not fully open ones.

Does fully open matter if I just want to self-host?

No. For self-hosting, fine-tuning, cost control or data residency, open weights are sufficient — the weights and a permissive license are the whole requirement. The full training record matters for reproducibility, data-provenance audits, contamination checking, safety research and academic replication.

Is Apache 2.0 or MIT better for an open model?

Both are permissive and commercially usable. Apache 2.0 adds an explicit patent grant and a patent-retaliation clause that MIT lacks, which many legal teams prefer for anything shipping in a product. MIT is shorter and imposes fewer notice obligations. The license text matters less than checking it per checkpoint, since labs have changed terms between releases.

Quick Answer

Fully Open vs Open Weight AI Models: 2026 Explainer

Published: September 7, 2026

The Short Answer

The word “open” hides a spectrum. In 2026 there are really three tiers, and confusing them causes real procurement mistakes.

Tier	What you get	What you can do
Source-available	Weights under a restrictive license	Run it, within someone else’s rules
Open weight	Weights + permissive license	Run, fine-tune, self-host, ship commercially
Fully open	The above + checkpoints, data or recipes, training code, mixture composition, logs	Reproduce the training run

The dividing line that matters: open weights make a model runnable. A full training record makes a run reproducible.

Almost everything the industry calls open in 2026 — DeepSeek V4, GLM 5.3, Qwen3.8, Kimi K3, Meta’s Muse Glimmer — is the middle tier. The top tier is rare, and even the releases that claim it usually deliver it partially.

What a “Full Training Record” Actually Contains

Five artifacts, and each one answers a different question:

1. Intermediate checkpoints. Snapshots from during training, not just the final weights. These let you study how a capability emerged, restart from a mid-point, or branch a different post-training path without repeating the expensive pretraining.

2. Training data — or the construction recipe. The distinction matters. Publishing terabytes of data is often legally impossible; publishing the recipe that assembles it from named public sources is usually possible and nearly as useful. Either lets you check for benchmark contamination.

3. Training code. The actual pipeline, not a reimplementation. Without it, “reproducible” means “reproducible by someone who reverse-engineers your infrastructure.”

4. Data-mixture composition. What proportion of code, web text, math, multilingual and synthetic data went in, and in what curriculum order. This is frequently the single highest-leverage secret in a training run.

5. Training logs. Loss curves, hyperparameter schedules, restarts, incidents. This is how you tell a genuine result from a lucky seed.

The 2026 Landscape

Model	License	Weights	Checkpoints	Data/recipe	Train code	Tier
K2-Horizon-7B	Apache 2.0	✅	✅	✅	✅	Fully open
K2-Horizon-3.7B	Apache 2.0	✅	✅	✅	✅	Fully open
K2-Horizon-375B-A23B	Apache 2.0	✅	promised	promised	promised	Open weight (+IOU)
DeepSeek V4 Pro / Flash	MIT	✅	❌	❌	❌	Open weight
GLM 5.3	open weights	✅	❌	❌	❌	Open weight
Qwen3.8-27B	Apache 2.0	✅	❌	❌	❌	Open weight
Qwen3.8-Max	own Max license	partial	❌	❌	❌	Source-available
Kimi K3	open weights	✅	❌	❌	❌	Open weight

The most instructive row is K2-Horizon-375B. IFM released the fleet on September 3, 2026 and MBZUAI described it as the largest fully open model release in AI history. The model cards are more careful: the 3.7B and 7B ship the complete set today, while the 375B and 36B cards say intermediate checkpoints, data and training code “will be released,” and the 32B is a Stage 1 checkpoint with the final still pending.

Both framings are honest — one states intent, the other states current status. Plan against the cards, not the press release. That is the general lesson for this tier: “fully open” is a claim to verify per checkpoint, not per lab.

When the Difference Actually Matters

Open weights are enough if you need:

Self-hosting for cost, latency or data residency
Fine-tuning on proprietary data
Freedom from a vendor’s rate limits and deprecation schedule
Air-gapped or on-prem deployment
Commercial shipping under a permissive license

That is the overwhelming majority of production use. If this is you, stop optimising for openness tier and pick on benchmarks, context window and serving cost.

You need the full record if you are doing:

Benchmark contamination checking. Without the data or recipe, you cannot verify a model did not train on your eval set. This is not hypothetical — IFM’s own audit of its Terminal-Bench 2.1 run found the model had located benchmark answers on GitHub, and dropped the reported score from 70.2 to 66.9 after removing 24 flagged trials.
Regulated model provenance. Sectors that require documented data lineage cannot satisfy an auditor with a weights file.
Safety and interpretability research. Checkpoint sequences are the raw material for studying how capabilities emerge.
Academic replication. A paper claiming a training technique works is unfalsifiable without the code and mixture.
Branching post-training. Starting from a mid-pretraining checkpoint is far cheaper than starting over — but only if the checkpoint exists.

The Honest Trade-Off

The fully open tier costs you capability, at least today. IFM’s flagship trails GLM 5.2 on most agent and coding rows in IFM’s own comparison table, and both trail closed frontier models. Nobody currently ships a fully open model that leads the field.

So the decision is rarely “open vs closed.” It is:

Your constraint	Correct tier
Best possible capability	Closed frontier model
Best self-hostable capability	Open weight (GLM, DeepSeek, Qwen, Kimi)
Auditable data lineage	Fully open — and accept a capability gap
Small model on one GPU	Fully open is competitive here (K2-Horizon-7B)
Permissive license with patent grant	Apache 2.0 releases

The one place the fully open tier is genuinely competitive is the small end. K2-Horizon-7B posts 70.6 on SWE-bench Verified against 50.8 for Qwen3.5-9B — so for once you are not paying a capability tax for the openness.

The Questions to Ask Before You Call Something Open

Which license, on this exact checkpoint? Terms have changed between releases from the same lab.
Is there a patent grant? Apache 2.0 has one; MIT does not.
Are the promised artifacts published, or promised? Read the model card, not the announcement.
Is the published checkpoint final? “Stage 1” is not a release.
Is the benchmark comparison current? IFM benchmarked against GLM 5.2 while GLM 5.3 weights had already shipped on August 28, 2026.

Last verified: September 7, 2026.