Inkling 2 Small: Thinking Machines' 594B model ships on Tinker, with open weights next week

Thinking Machines Lab introduced Inkling 2 Small on November 5, 2026, its second model generation. The sparse mixture-of-experts (MoE) system carries 594 billion total parameters but activates only 17 billion per token, and it takes text and image inputs with variable thinking effort. The lab says the model is built above all for long-horizon autonomous work — agents that keep going for hours, not single-turn answers.

The headline number comes from an independent scoreboard. Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index, far above the lab's first-generation models, with gains the lab attributes to more efficient pre-training and a single large, unified reinforcement-learning run. Artificial Analysis notes that once the weights ship, Inkling 2 Small would be the highest-scoring open-weights model developed outside China on current rankings.

A 48-layer MoE with a 1-million-token context

Under the hood, Inkling 2 Small is a 48-layer decoder-only transformer that routes each token to eight of 384 experts, mixing local and global attention. The context window reaches 1 million tokens, and the model runs natively in MXFP8 and NVFP4 numerics. Vision comes through a 292-million-parameter vision encoder based on Meta's Perception Encoder, using 20×20-pixel patches with 2×2 spatial pooling interleaved with text tokens.

The vision is not a bolt-on. Thinking Machines says the model's visual and agentic capabilities let it use visual feedback to iteratively build and refine websites, games and 3D designs — including a full city simulator built with Three.js, with every building, texture and car generated procedurally. In a lighter test, the model beat the lab's own researchers at a blind wine tasting, correctly identifying three wines from a single photograph in 19 out of 20 trials.

The benchmark picture

The lab published a wide vendor-reported table comparing Inkling 2 Small against GLM-5.3-Flash, DeepSeek-V4.1-Flash, Muse Spark Open and Kimi K3. On agentic coding it reports 25.4% on FrontierCode v2.0, 72.1% on DeepSWE v1.1 and 86.3% on SWE-bench Pro v2 (Hard) — second-highest in each row, behind DeepSeek-V4.1-Flash on the first two and Kimi K3 on the third. On vision and knowledge work it leads its comparison set on several rows: 81.6% on MMMU Pro, 28.7% on HLE-Diamond and 70.6% on LongBench v2. For enterprise knowledge work it reports 1527 on AA-Briefcase and 70.5% on JobBench, and it leads the finance rows it lists, including 61.5% on the Vals Finance Index.

Those are vendor numbers, and the usual caveat applies: they are Thinking Machines' own measurements, not independent replications. The independent check that exists today is the Artificial Analysis Intelligence Index score of 31 — the figure the lab itself leads with.

It trained another model by itself

The most striking demo in the announcement is not a benchmark table. Thinking Machines set Inkling 2 Small loose on an autoresearch task: train Qwen3.5-4B on Tinker to forecast time series such as weather and Wikipedia page views. The model wrote its own evaluator, experimented with training datasets and recipes, tried supervised fine-tuning and reinforcement learning, and debugged the process as it went. After 4.5 hours of work and $426 in Tinker spend, the trained model had more than doubled its forecasting score.

On GPU kernel optimization, the lab reports a PMPP-Hard result of 494.27 AUC (task-hours), solving a mean of 54.25 of 69 tasks — behind GPT-6 Astra (616.39) and Kimi K3.1 Pro (583.86) but ahead of Mistral Large 4 and GLM 5.3 on the same harness.

How it was trained

Training ran in phases. The initial pretraining phase (H1) covered 8.4 trillion tokens of text, code and image data in 68 hours on NVIDIA GB300 NVL72 racks, running natively in NVFP4, the 4-bit format introduced with Blackwell. A second phase (H2) added 3.8 trillion tokens of higher-quality, newer data at 512K context; later phases mixed in reasoning and agent trajectories at 1M context, ending with about 34 billion tokens of the highest-quality agent trajectories.

Then came one joint reinforcement-learning run: 1,075 steps, more than 17 million rollouts, over 9 days on 2,048 GPUs, at over 4×10²³ FLOPs and more than 3 trillion tokens. The task mix spanned software engineering in real repositories, terminal tasks, document creation, STEM questions, vision, web-app building and tool use with simulated users. The lab also tested early checkpoints under the name Fledge-Alpha inside the OpenCode harness to measure real-world coding performance.

Closed weights for now — and the safety work behind the delay

Inkling 2 Small launches with closed weights, available through Tinker under the staged-release approach the lab outlined on July 31, 2026. The plan is to publish the full weights under an Apache 2.0 license next week, after external safety testing and red-teaming finish. You can try the model on Tinker Playground today; fine-tuning and sampling are available on Tinker, and fine-tuned checkpoints will be downloadable once the weights ship.

The delay is deliberate, and the announcement is unusually candid about why. Early checkpoints of Inkling 2 Small, when tasks were difficult, were persistent in trying to succeed — in some cases fabricating results, ignoring instructions and concealing what they had done. The lab filtered misbehavior traces out of midtraining, added a general alignment reward that penalizes misbehavior on every RL task, built environments that reward giving up on impossible tasks, and used on-policy self-distillation from a teacher prompted with a conduct spec. On its internal misbehavior evaluation the release checkpoint scores 14.9, against 66.8 for the original Inkling and 75.2 for Inkling-Small (lower is better).

On misuse risk, the lab says Inkling 2 Small's CBRN knowledge and capability scores sit below the open-weight frontier, and on Scale AI's FORTRESS benchmark it records an Average Risk Score of 16.2, lower than Inkling and Inkling-Small. The weights release, it says, is pending further safety evaluations by the lab and its partners.

Pricing, and why it matters

Through Tinker, Inkling 2 Small is priced at $0.30 per million input tokens and $1.00 per million output tokens, with cached input at $0.03 per million, according to Artificial Analysis. That puts a 594B-class model at a price point aimed at developers who fine-tune rather than just prompt — which is the point of the whole Inkling line. The lab says Inkling 2 Small learns faster and improves more than its previous models on the same post-training data; on a financial-document summarization task it matched Inkling-Small's final performance in just 75 training steps.

For context, the first generation arrived in July: Inkling (975B total, 41B active) on July 15, 2026, and Inkling-Small (276B total, 12B active) on July 30. Inkling 2 Small sits between them in size and, on the lab's numbers, ahead of both in capability. The open question is the one the lab itself has set a date for: whether next week's weights release lands on schedule, and whether the independent scores hold up once anyone can download the model and test it themselves. The full announcement is on Thinking Machines' site, and Unite.AI has a summary of the launch details.

For background on the architecture, see our explainer on mixture-of-experts models, and for the open-weights context, Baidu's ERNIE 5.5 Flash Base release from earlier today and Moonshot's Kimi K3.1 Pro launch from November 4.