Weights pending
Two model releases landed on November 5, 2026, and together they make a neat pair, because they sit at opposite ends of the same bargain. Thinking Machines Lab announced Inkling 2 Small — 594 billion total parameters, 17 billion active, a sparse mixture-of-experts model aimed at coding, visual work, and long-horizon autonomous agents — and shipped it with closed weights, promising the full weights under Apache 2.0 "next week," once external safety testing and red-teaming are done. The same day, Baidu released ERNIE 5.5 Flash Base — 280 billion parameters, 14 billion active — and put the raw pre-training checkpoint straight on Hugging Face, months before any instruction-tuned version exists.
One company has a finished model and no weights. The other has weights and no finished model. Both are trading on the word open, and both make you wait for something. The question this site exists to ask applies to both: but is it open?
The book runs every release through the same five questions — Access, Permission, Deployment, Transparency, Reproducibility. Let's run both models through it.
ERNIE 5.5 Flash Base first, because you can actually download it.
- Access: yes. FP8 and BF16 safetensors are on Hugging Face and ModelScope. This is the strongest "yes" of the two, and it is unqualified — the checkpoint is there, all of it.
- Permission: read the license. This is not MIT and it is not Apache 2.0. The weights ship under the ERNIE 5.5 Flash Base Model License Agreement v1.0, a custom document with a copyright line reading "All Rights Reserved." Whatever that agreement permits, it permits by Baidu's grace, not by a standard anyone else wrote. A custom license can be generous; it can also change in v1.1. The openness here is real today and revocable in principle.
- Deployment: yes. Baidu documents serving with vLLM and SGLang, up to a 1M-token context. This is a model you can run on your own hardware, which is the heart of the practical test: can you escape the hosted API? You can.
- Transparency: partial. There is a technical blog, three papers (on the attention design, the router, and the training trick called ECHO), and a headline number — 33.5 trillion tokens of pre-training data. What there is not: a dataset composition, a filtering account, or a description of how the corpus was assembled. You know how big the meal was, not what was in it.
- Reproducibility: no. A base checkpoint is not a recipe. Nobody is retraining this model from the artifacts provided.
That lands ERNIE at roughly Tier 3 on the book's spectrum — practical open-weight. Strong on Access, Deployment, and (today) Permission; partial on Transparency; absent on Reproducibility. A serious release, and genuinely more open than most of what calls itself open — with the caveat that the "more open than most" rests on a license only one company wrote.
Now Inkling 2 Small.
- Access: no. At launch the model lives on Tinker, Thinking Machines' own platform. Playground gives inference; fine-tuning happens under platform terms. No weights to download.
- Permission: promised. Apache 2.0 is the stated destination, which would be a genuinely permissive license — but a promise is not a license. Nothing is granted until the weights exist.
- Deployment: pending. Follows the weights.
- Transparency: thin. A blog post describes the safety evaluations — chemical, biological, radiological and nuclear risk, cyber, refusal-suppression, agentic capability — but the training process itself is largely a closed book.
- Reproducibility: no.
So Inkling sits at Tier 1 at launch: access-limited, source-visible only through a hosted service, with a stated intention to reach Tier 3 within a week or two.
Two observations, because the pair is more interesting than either release alone.
The first is the governance-shaped delay. Thinking Machines published a post called "a safe path to open weights," and Inkling 2 Small is the first model shipped under it: capability first, weights after testing. The delay itself is not new — companies have sat on weights forever. What is new is making the delay an announced, scheduled, publishable part of the release, like a software release candidate. That is a governance story wearing a shipping schedule, and it is worth watching where the line settles. A one-week delay is a press release; a one-year delay is a decision about who gets to run frontier models at all. The book's chapter on managing cost and capacity cares about this too, because weights that arrive late arrive to a market that has already organized around the hosted API.
The second is base-before-instruct. Baidu's stated reason for releasing at the end of pre-training is that different domains, reasoning tasks, and agentic applications need different data and different behavioral policies — so they hand you the general capabilities and let you decide where the model goes next. That is a different flavor of openness than a polished chat model under a permissive license: it is openness to shape the thing before anyone has decided what it should say. It is also, frankly, less usable for most people — you get a completions endpoint, not an assistant, and the instruct version "will follow." But as a statement about who controls a model's character, it is the more radical of the two releases.
So — but is it open?
ERNIE 5.5 Flash Base: mostly. Downloadable, runnable on your own hardware, under a custom license you should actually read, with the training account left unwritten. Inkling 2 Small: not yet. The model exists, the benchmarks are public, and the open part is scheduled — which means that for one week at least, "open" is doing the work of a delivery date. Ask again when the weights land, and check the license they land under.
Adapted from the draft manuscript of But Is It Open?