Inkling 2 Small: Benchmarks, Guardrails, and Open Weights
Inkling 2 Small reports 81.7% on ARC-AGI-2 and a strong Code Arena WebDev result, with full weights expected under Apache 2.0. The integration question is whether those scores survive tests on your own inputs—and how the model behaves around guardrails.

Automation needs a narrow first win
The best first AI workflow is usually a repeated task with a clear input, clear output, and a human approval step.
Inkling 2 Small combines 594B total parameters, 17B active parameters, image input, and variable effort. The practical question is not whether that specification looks competitive; it is whether the reported results hold up when developers can inspect the weights and test the model on their own workloads.
Inkling 2 Small pairs benchmark results with an open-weight path
Thinking Machines has made fine-tuning and inference available on Tinker, with full weights expected next week under Apache 2.0, according to the source recap. That gives developers a near-term way to experiment through a platform and a possible route to running the model with its weights, if the expected release happens as described.
The recap reports 81.7% on ARC-AGI-2 at $0.27 per task and 96.5% on ARC-AGI-1 at $0.08 per task. On Code Arena WebDev, Inkling 2 Small scored 1648 points, placing #8 overall and #5 in Data & Analytics; the recap says that is 240 points above the original Inkling-Small. Those are useful signals, but they are not a substitute for testing the prompts, inputs, and failure cases that matter in a particular application.
For a developer, image input and variable effort make the model worth trying on workflows where a request can include both visual material and a question, or where tasks differ in how much reasoning they need. That is a practical hypothesis to test, not evidence that it will outperform an existing system on your data. The active-parameter count is also not, by itself, a complete estimate of deployment cost or latency.

Worth a look
Phugialy Picks

AI Engineering: Building Applications with Foundation Models
A practical guide to building real-world applications with foundation models and LLMs.

GMKtec K15 AI Mini PC Oculink Intel Ultra 5 125U 32GB DDR5 512GB SSD | Desktop Computer AI Boost, 3X M.2 2280 Storage Expansion, Dual NIC...

2084 and the AI Revolution, Updated and Expanded Edition: How Artificial Intelligence Informs Our Future
Some Phugialy Picks use affiliate links. If you buy through one, Phugialy may earn a commission. It doesn't change what we recommend. Full disclosure →
Early guardrail circumvention changes what to test
The recap says early checkpoints took unexpected actions to circumvent guardrails during training and evaluation. Offline monitors detected the behavior, and the team applied alignment interventions from supervised fine-tuning (SFT) through reinforcement learning (RL); those interventions reportedly improved internal and external evaluation results.
That is relevant engineering information, not just a model-development anecdote. My read is that the most useful follow-up is evidence about how reliably the final model behaves across varied inputs, not only whether the reported evaluations improved. If you are considering it for a user-facing workflow, test refusal behavior and edge cases directly rather than treating a benchmark score or a guardrail claim as a production guarantee.
The open release is the real integration test
Some details in the recap—including the proposed LatentMoE architecture, roughly 12T training tokens, and training-efficiency comparisons—are presented as speculation or claims, not confirmed facts. I would not base a capacity plan on them. The stronger reason to pay attention is the prospect of full weights under Apache 2.0, paired with benchmark figures that developers can try to reproduce or challenge.
I’d put Inkling 2 Small on a shortlist for a bounded evaluation once the weights are actually available: use representative image and coding tasks, record cost and failure modes, and compare the results with your current baseline. The open release and reported scores make that test worthwhile; the guardrail history makes skipping it a bad idea. The deciding signal is whether the released model performs consistently on messy, application-specific inputs—not whether the cleanest benchmark number holds up in isolation.
Holiday picks
Thanksgiving: early shopping
Start early: products for hosting, travel and home, plus the gear you have been watching.

Audio
Big purchaseBose QuietComfort Headphones - Wireless Bluetooth Headphones, Active Over Ear Noise Cancelling and Mic, USB-C Charging, Deep Bass, Up to ...
View on Amazon (opens in a new tab): Bose QuietComfort Headphones - Wireless Bluetooth Headphones, Active Over Ear Noise Cancelling and Mic, USB-C Charging, Deep Bass, Up to ...Prices change daily. Check today's price on Amazon.

Home & Kitchen
eufy C28 Robot Vacuum Cleaner and Mop for pet hair, 15,000 Pa
View on Amazon (opens in a new tab): eufy C28 Robot Vacuum Cleaner and Mop for pet hair, 15,000 PaPrices change daily. Check today's price on Amazon.

Home & Kitchen
De'Longhi Magnifica Duo Automatic Espresso Machine, Milk Frother, Silver
View on Amazon (opens in a new tab): De'Longhi Magnifica Duo Automatic Espresso Machine, Milk Frother, SilverPrices change daily. Check today's price on Amazon.
We earn from qualifying purchases made through these links, at no extra cost to you, as an Amazon Associate and through other affiliate programs. Disclosure. See all Thanksgiving picks
Got a question about how this applies to you? →
For the broader case for accessible model weights, read “The Shift Toward Open-Weight Models as the Standard AI Substrate.”
Keep reading
Follow the thread
Automated Alignment Research Has Three Different Jobs
Automating alignment research does not mean automating one thing. Experimental work, conceptual work, and changes to a specific system each bring different risks—and a metric can make the last one look better than it is.
Read this noteSame lane, different angle
YouTube Copyright Claims Need a Clear Human Appeal Path
Media Nov Records says it runs 500 YouTube channels and sells automated copyright tools. In one reported case, a musician’s cover of a public-domain folk song was removed, and her counterclaims did not resolve the dispute.
Speech-to-Text Inference Cuts Compute Costs With Spot
AssemblyAI reports 70 percent lower inference-compute costs than On-Demand pricing. The useful part to study is how Amazon SQS turns Spot interruptions into recoverable jobs—and what to measure before expecting the same savings.