Thinking Machines drops Inkling 2 Small, a massive MoE coder with vision

Most open-source models fall apart the moment you ask them to run a realistic, multi-step coding agent loop.

Thinking Machines wants to change that. On November 5, the company introduced Inkling 2 Small, a massive Mixture-of-Experts model packing 594B total parameters with just 17B active during inference.

It scored 31 on the Artificial Analysis Intelligence Index, comfortably clearing its predecessors. While weights are temporarily gated on Tinker Playground for external red-teaming, full Apache 2.0 open weights are scheduled to drop next week.

Eyes inside the coding loop

Inkling 2 Small's defining trick is how it weaves multimodal perception directly into its developer workflow.

Equipped with a 292M-parameter vision encoder, the model processes interleaved image and text tokens. This lets it inspect rendered visual outputs while building websites, games, and 3D scenes.

Instead of blindly editing code, it looks at the rendered UI, spots alignment flaws, and iterates autonomously. In one tongue-in-cheek internal test, the model even beat researchers at visual wine tasting, identifying three bottles from a single photo in 19 out of 20 rounds.

On the coding side, it was trained using reinforcement learning across hundreds of thousands of long-horizon tasks. An agentic RL grader evaluated its step-by-step problem-solving strategy rather than relying strictly on binary unit-test passes.

Early checkpoints tested under the alias Fledge-Alpha confirmed that the model follows nuanced system prompts closely and habitually double-checks its edits without being prompted.

Training other models for $426

Beyond building web interfaces, Thinking Machines pushed the model toward autonomous scientific research and GPU kernel engineering.

In one end-to-end experiment, Inkling 2 Small was tasked with training a smaller Qwen3.5-4B model on Tinker to forecast weather and Wikipedia traffic trends.

The model wrote its own evaluation harness, tested different dataset mixtures, experimented with SFT and RL pipelines, and resolved bugs along the way. After 4.5 hours and $426 in compute spend, the trained model more than doubled its baseline forecasting score.

It put up similar fight on the brutal PMPP-Hard kernel optimization benchmark, matching high-tier proprietary systems with a 64 out of 69 pass@4 record.

Why it matters

If you build autonomous software agents or complex UI workflows, Inkling 2 Small is one of the most promising open releases this quarter.

Most existing coding models generate code well in single turns but struggle when left unattended inside an IDE or terminal. Giving the agent eyes to inspect visual outputs directly solves one of the biggest friction points in generative web development.

Releasing a 594B MoE under Apache 2.0 also heaps pressure on frontier labs charging high API rents for agentic reasoning. With only 17B active parameters, per-token inference remains relatively fast.

That said, we think hardware logistics will still keep hobbyists at bay. Storing 594B parameters in memory requires serious enterprise VRAM, regardless of how few experts fire simultaneously.

Furthermore, letting an agent burn $426 over four hours to train another network shows what is possible, but it also highlights that autonomous problem solving can get expensive fast if left without guardrails.