Parameter Update: 2026-27

Surprisingly big week for model drops! Really thought Anthropic would have a little bit longer before China catches up to Fable, but shows what I know, I guess.

Thinking Machines Inkling

The least performant, but most interesting model drop of the week comes from Thinking Machines. Inkling is a ~1T parameter, 1M context model trained to be natively multimodal, i.e., it was trained on a mix of text, images, audio and video and can reason across the former three directly. In pure text performance, it falls squarly into the Opus 4.8 class, that has become increasingly crowded over the last couple weeks.

There are a few reasons Inkling is interesting. The main one, in my mind, is that is doubles down on Thinking Machines (admittedly short) history of multi-modality + fine-tuning as differentiators - it scores #1 in some audio benchmarks, and is directly available on Tinker. This, combined with the open weights release is not something we see a lot of these days, at least not from western labs.

The second is that it appears to be one of the few big pretrains we still see that haven't been bootstrapped on large amounts of Claude data (though Thinking Machines has admitted to fine-tuning on some Kimi K2.5 traces).

The final thing to look at is the fact that their RL curves don't appear to be plateauing, which means that can probably expect more powerful models very soon.

Kimi K3

The second model drop of the week is Kimi K3. At 2.8T parameters, it's a much bigger model. Given the extended K2.5 series was already close to Opus, it should be no huge surprise (but it somehow still is), that K3 reaches Fable/Sol level performance in some tasks.

The model will go open weights on 27.07. - until then, you can use the model directly in the Moonshot/Kimi stack with API pricing around Sonnet (i.e., 3x cheaper than Fable). Still pricy for a Chinese model, but significantly nicer for day-to-day use than Fable.

With the Kimi context, it's interesting to look back at the Fable drama over the last couple weeks. News of Mythos Preview first came up in April, and we discussed if Anthropic would ever release a model of this capability tier. It's now July, and you can pay a Chinese lab $19 to get it as an open-weight coding agent that's so popular, they've had to temporarily pause signups

Qwen3.8

Not content to let Moonshot have all the glory, the Qwen team decided to pre-release Qwen3.8 the same day. We don't actually have many benchmarks for it yet (they also appear to keep swapping model checkpoints?), but the model is usable in their token plan now, with open weights coming soon. Similarly to Kimi K3, performance seems neck-in-neck with Fable 5.

Anthropic Fable Access

Given the news above, it's unsurprising Anthropic decided to keep offering Fable 5 in Claude Code max plans (though it is getting rid of it entirely on the $20 plan). Still only getting up to 50% of included usage, at the insanely high API pricing, makes the whole thing seem like a questionable deal though.

Hugging Face Agentic Cyberattack Incident

Over the last couple months, we've talked a lot about the increased potential for autonomous cyberattacks driven by all the new models. This week, we saw a concrete example of what these attacks might look like in practice. On Thursday, HuggingFace reported a security incident, where an attacker exploited two issues in their dataset processing pipeline to harvest credentials and move into their infrastucture. The impact of the attack has been hard to quantify, as the agent executed "many thousands of individual actions across a swarm of short-lived sandboxes, with self-migrating command-and-control staged on public services".

While impact appears to be limited for this specific attack (there is no sign the attacker managed to access user data or compromise supply chains), there are valuable lessons to be learned from it. For instance, the majority of the defensive work they did in the aftermath was also driven by AI. Funnily enough, then, the main issue HuggingFace reports on that end is that their own models would sometimes refuse to conduct some of the security and analysis work for safety reasons. They also note practical problems with getting large datasets into proprietary models.