Introducing Inkling 2 Small

Pippa They just dropped Inkling 2 Small, and the wild part isn’t the 594 billion parameters, it’s that this thing is basically built to run agents for hours at a time.

Tyler Yeah, and they’re promising Apache 2.0 weights next week after safety review, so it’s not just another API you poke once and forget.

Pippa So quick sketch: it’s a Mixture-of-Experts model, five hundred ninety four billion total parameters but only seventeen billion “active” per token, multimodal so it can take images, and it has that variable effort knob for more or less compute per request.

Tyler Right. MoE in plain terms is: you’ve got a bunch of specialist subnetworks, and for each token the router picks a couple of them instead of running the whole giant thing. So you get big total capacity, but inference cost closer to a medium model.

Pippa Which is exactly the sweet spot if you’re running an agent that’s rewriting codebases or 3D scenes for ten hours, not just answering a chat question.

Tyler They’re leaning into that. The headline is it scores thirty one on the Artificial Analysis Intelligence Index, which is their composite across coding, vision, reasoning, knowledge work. But the parts that matter to me are the long-horizon coding and autoresearch sections.

Pippa Yeah, talk through that, because that’s where this stops being a big number and starts sounding like a product story.

Tyler So for coding, they did reinforcement learning on hundreds of thousands of long-horizon tasks. And instead of just “did your code pass the unit tests”, they used what they call agentic RL graders. Those look at how the model approached the problem: did it plan, did it refactor, did it run tests, that sort of thing.

Pippa Mm-hm.

Tyler They also trained across multiple agent harnesses and context-compaction strategies, which is nerd-speak for: they didn’t overfit to one orchestrator, they tried to make it robust to different ways of chunking and feeding code. That’s the bit that makes me think it might actually behave in the messy real world.

Pippa And they did the thing I always want but rarely see: they tested early checkpoints on OpenCode, under this Fledge-Alpha alias, to see if it could handle real coding tasks instead of just synthetic ones.

Tyler Exactly. On the classic benchmarks: DeepSWE v1 point one, SWEBench Pro version two Hard, Vibe Code Bench, it’s usually sitting second in the row. DeepSeek V4 point one Flash or Kimi K3.1 Pro still edge it out, but Inkling 2 Small beats GLM five point three Flash and Muse Spark Open by a decent margin on most of those.

Pippa So if you’re a team that wants open weights, that’s… kind of the point, right? Once those Apache weights land, you’re in “near frontier coding” territory without having to call a closed API.

Tyler With the usual “once those weights land” asterisk, yeah.

Pippa Yeah, yeah, we have a whole scoreboard of people slipping a week on open-weight promises. I remember your Kimi prediction faceplant.

Tyler I earned that one. So I’m not betting dates here, but at least the external trackers you and I read are already watching whether the Apache drop actually happens next week.

Pippa The kernel-optimization bit really jumped out at me. That PMPP-Hard benchmark where each model gets ten hours to speed up GPU kernels from TileLang—how did Inkling do there?

Tyler So PMPP-Hard is sixty nine GPU kernel tasks, four independent attempts, ten-hour budget. They rank models by area under the “tasks solved over time” curve. GPT-6 Astra is on top, then Kimi K3.1 Pro and Claude Opus five point five. Inkling 2 Small is in that next cluster: about fifty four point two five tasks solved by ten hours, with an area-under-curve just shy of DeepSeek V4.1 Flash.

Pippa So not best-in-class, but in the same band as the big closed names, on a benchmark that is literally “can your agent grind on annoying kernel bugs for a workday.”

Tyler Yeah. And that’s with all models using their max reasoning setting, the same PI harness and a shared compaction policy. So at least on paper, you’re comparing similar agent setups.

Pippa This is where I kind of want to wave the open-versus-closed flag for a second. You’ve got Kimi K3.1 Pro, DeepSeek V4.1 Flash, Claude Opus, GPT-6 Astra at the top of these long-horizon and kernel things. If Inkling 2 Small really ships Apache weights and keeps this behavior, that’s a different tier of open option than “pretty good chat model you can download.”

Tyler Yeah, it’s closer to “you can self-host something that was actually tuned for ten-hour agents,” not just for one-shot prompts. That’s new.

Pippa And then there’s the autoresearch demo. Inkling 2 Small basically sits on Tinker and trains Qwen 3.5 four B to do time-series forecasting for four and a half hours with a five hundred dollar budget cap, and ends up more than doubling the forecasting score.

Tyler That walk-through is surprisingly concrete. It inspects the Tinker usage, writes a little eval script, runs a baseline—gets skill around zero point one on a few windows—builds a training set of twelve thousand five hundred windows, runs S F T as a smoke test, catches a bug where all eight samples are identical because it reused a seed, fixes the eval, then does full S F T and considers an RL phase.

Pippa And it keeps an eye on spend. There’s the aside like “evals are expensive, we’ve already spent about one hundred and one dollars,” and it checks batched eval cost at two cents per window. That’s the kind of behavior you want from an agent that’s allowed to touch your cloud bill.

Tyler Totally. It’s still a cherry-picked success story, but it shows the loop we all talk about: define the task, build your own grader, iterate on data, debug your own training script. It’s not just “I wrote one bash command and magic happened.”

Pippa Okay, so if you’re actually building with this, what’s the move once weights show up? Because right now it’s Tinker-only, you can hit the playground, fine-tune, sample, but the fun starts when you can run it on your own hardware.

Tyler If I had an existing long-running coding agent, my first test would be: swap the model to Inkling 2 Small in the same harness and see if the “follows instructions better, verifies its work without being asked” claim actually shows up. Like, does it run tests unprompted, does it think before rewriting huge files.

Pippa And I’d steal their homework on kernel stuff. Take a couple of your own slow kernels, or even pull the public TileLang ones they mention, and see if an Inkling-powered agent can get you meaningful speedups within some budget. If that works, that’s a pretty direct productivity story for infra teams.

Tyler The brave version is trying a tiny replica of that time-series forecasting run. Use a much smaller dataset and budget, let it design the evaluator and training loop, and watch where it stumbles. If it can reliably get from “base model skill around zero point one” to “hey, we actually improved this metric” on your data, that’s a strong signal the autoresearch thing isn’t just a one-off blog demo.

Pippa Yeah, that’s the part I’m honestly excited about. Not that it scores thirty one on some index, but that someone might finally be shipping an open model whose whole personality is “I can work on this for half a day without you babysitting me.”

Tyler As long as that Apache zip file actually appears next week.

Pippa We’ll be refreshing together, Tyler. For now, I’m going to go poke the Tinker playground and see how far I can get before I hit the budget alerts.