Ali Asif

The AI log, open weights · frontier · science

Inkling 2 Mini, OpenAI on distillation, MedChemBench

Thinking Machines opens a 26B MoE with a looped decoder, OpenAI details its anti-distillation stack, plus MedChemBench and a self-rewriting weights paper.

By , published

What shipped

  1. Thinking Machines

    Welcome Inkling 2 Mini

    A multimodal MoE with about 26B total and 3B active parameters, a 1M token context and a looped middle of the decoder, with day-zero support in Transformers, llama.cpp, MLX and SGLang.

  2. OpenAI

    Combatting adversarial distillation

    Reports reasoning-extraction attack success falling from 54.0% to 0.2% on the newer model, and describes classifiers, API changes and account-level enforcement.

  3. OpenAI

    Can AI help scientists identify promising drug candidates?

    MedChemBench has 99 problems from 117 papers and patents; the best model scores 84.5% at xhigh reasoning, but selection and numerical prediction sit near half credit.

  4. arXiv

    Language Models Can Rewrite Their Own Weights

    A 27B open model edits its own weights without gradients and gains 6.6 points on unseen problems, per the abstract.

  5. ElevenLabs

    Video variants and localization in Ads Engine

    Ads Engine connects to Google, Meta, LinkedIn and TikTok accounts and produces new hooks, soundtracks and dubbed versions of existing video ads.

  6. mankidev

    antialgorithm

    An MIT-licensed agent skill that makes the model name its default pick for creative work, then choose a different one for a stated reason.

A quiet-ish Friday with one open model worth downloading and one lab explaining how it protects its reasoning. Everything below links the primary page I read.

Inkling 2 Mini

Thinking Machines released Inkling 2 Mini, a multimodal mixture-of-experts model with about 26B total parameters and 3B active per token. The router picks 8 of 128 experts per layer, the context window is 1M tokens, and it takes images and text.

The odd part is the decoder. There are 24 physical layers, but layers 6 to 17 run three times in a loop, so each token sees 48 logical layer visits. A research variant supports one to three loop passes. Attention mixes sliding-window layers (1,024 tokens) with a full-attention layer in every block of six.

The post says it leads 32 of 48 benchmarks against Gemma 4 31B, Qwen3.8-27B and a few others, but the numeric scores live in the model card, so I can't repeat them here. Support is there from day zero: Transformers, llama.cpp with GGUF quants, MLX and SGLang. The speculative decoding numbers were measured on a single B300.

OpenAI on adversarial distillation

OpenAI describes how it detects and disrupts people training a second model on its outputs. The measures are adversarial training so models refuse to hand over protected reasoning, input and output classifiers, API changes that tie reasoning to the conversation that produced it, and enforcement against coordinated account networks.

The headline number: recovering reasoning from an earlier request succeeded 54.0% of the time against GPT-6 Astra with the original safeguards and 0.2% against GPT-6.2 Astra with the updated ones. That is adaptive black-box red-teaming run by OpenAI, so treat it as a claim about their own tests.

MedChemBench

A new OpenAI benchmark for medicinal chemistry: 99 problems drawn from 117 papers and patents, covering 761 molecular structures. GPT-6.2 Astra scores 84.5% at xhigh reasoning. Structure reconstruction is strong, while compound selection and numerical prediction stay around half credit even for the top model. The post itself says it cannot show whether models discover new drugs, and that exposure to the source papers can't be ruled out.

Smaller items

The self-rewriting paper claims a 27B open model can diagnose its failures and edit its own weights with no gradients, gaining 6.6 points on unseen problems and restoring a damaged copy to over 80% of original accuracy on image tasks. The abstract does not name the benchmarks.

ElevenLabs Ads Engine pulls existing ads from Google, Meta, LinkedIn and TikTok, then writes new hooks, generates music and dubs each ad with lip-sync. Languages and pricing are not listed.

antialgorithm is a small MIT-licensed skill with helper scripts. The model states its default answer, then picks something else for a reason tied to the project.

What I'd watch next

Independent scores for Inkling 2 Mini once the community runs it on its own evals, and whether the looped layers hold up at one pass versus three. I'd also watch whether other labs publish their distillation defenses with the same level of detail.

Sources

Topics: OpenAI, arXiv, ElevenLabs, Inkling, Agent Skills, Distillation, Thinking Machines, MOE