Thinking Machines Lab released Inkling this week, a 975 billion parameter mixture-of-experts model, adding another credible name to a field that already includes Moonshot's Kimi, DeepSeek, Alibaba's Qwen, Zhipu's GLM and MiniMax. For Australian IT leads trying to keep a mental map of which open model is worth watching, the honest update is that the map keeps growing faster than most internal teams can track.
A new name in open weights
Inkling is Thinking Machines Lab's first major public release, and on paper the specs are impressive: a 975 billion parameter mixture-of-experts architecture, benchmark scores that sit near the top of several public leaderboards, and an open licence that lets businesses self-host or run it through third-party inference providers. It arrives roughly eighteen months after a wave of Chinese and US labs pushed open-weight models from a niche research curiosity into a genuine procurement category boards now ask about.
For a Melbourne fintech or a Brisbane logistics operator evaluating AI vendors this quarter, the practical question isn't whether Inkling is technically capable. It almost certainly is. The question is whether adding it to the evaluation shortlist changes the underlying decision, and for most businesses it doesn't.
Why another entrant matters less than it sounds
New labs entering the open-weight race is, on one level, good news: more competition tends to push managed inference prices down and quality up across the board over time. But for a business in Sydney trying to make a single AI infrastructure decision this quarter, a new 975 billion parameter model changes almost nothing about the fundamentals of that decision.
Every new open-weight release needs independent verification before its benchmark claims can be trusted for business use, and that verification takes weeks, not the days most social media hype cycles allow for.
A model from a first-time lab carries more operational risk (no track record on model updates, deprecation policy, or support responsiveness) than one from a lab with two or three years of shipping history.
The list of open-weight models a small Australian business could plausibly evaluate is now well past 50, which is itself a signal that most businesses should stop trying to pick the single best one and instead pick a stable primary vendor with a credible fallback.
What a credible open-weight lab actually looks like
Not every open-weight release deserves equal scrutiny, and not every one deserves equal dismissal either. Before we recommend a client even trial an open-weight model as a fallback, we look for a handful of concrete signals rather than benchmark scores alone.
A published deprecation and versioning policy, so a production workflow doesn't break without warning when the lab ships v2.
At least one prior model release with a visible track record of bug fixes, safety patches, and honest incident disclosure.
Transparent, published inference pricing across at least two hosting providers, not a single exclusive deal that can be renegotiated against you later.
Independent third-party evaluation results, not just the lab's own benchmark chart, ideally covering the specific task type you plan to use it for.
Inkling clears some of these bars and not others yet, which is exactly why the right move this week is to note it, not adopt it.
The cost of chasing every release
We've seen Australian businesses spend upward of $20,000 in internal engineering time evaluating a rotating cast of open models before ever shipping a single production workflow. That's money spent on model tourism, not on the business problem the AI was meant to solve, and it's a pattern we flag early in almost every AI strategy engagement we run.
Our recommendation for most SMB and mid-market clients is deliberately boring:
Pick one primary model, Claude for the governance, support and Australian data-handling reasons we've covered elsewhere, and one open-weight fallback for cost-sensitive, low-stakes tasks.
Revisit that choice on a quarterly cadence, not every time a new lab makes a launch announcement.
Budget evaluation time explicitly, so checking out the new model of the month doesn't quietly consume a sprint every few weeks.
How this fits into a Claude-first stack
Most of our Sydney and Melbourne clients run Claude as the primary model for anything customer-facing, financial, or subject to APRA or Privacy Act obligations, and reserve open-weight models for internal, low-stakes tasks where the cost difference is worth the extra operational overhead of self-hosting or a smaller inference provider. That split isn't brand loyalty, it's a governance decision: Claude comes with a documented support path, predictable deprecation timelines, and an enterprise agreement your compliance team can actually point to. A new open-weight model, however capable, doesn't replace that paper trail overnight.
Where this leaves you
Inkling may turn out to be a genuinely strong model. It may also be forgotten within a quarter, as several 2025 open-weight launches already have been. Either way, the discipline of not chasing every release is worth more to most Australian businesses than the marginal capability gain from switching this month.
If your team is stuck in an evaluation loop and wants a structured way out, book time with us and we'll help you set a model policy that doesn't need revisiting every time a new lab makes an announcement.
