Thinking Machines Lab unveils Inkling 2 Small: a 594B-parameter MoE with vision, an agentic focus and planned open weights

Note: the text received is cut off near the end, during the wine-tasting example. The sections on personalization, reasoning and epistemics, safety and alignment, training, benchmarking and limitations were not received. This summary is based only on what appears in the material.

Thinking Machines Lab has introduced Inkling 2 Small, the first model of its second generation. It is a Mixture-of-Experts model with 594 billion total parameters and 17 billion active. It supports image inputs and variable effort. The company says it clearly outperforms its previous models on the Artificial Analysis Intelligence Index, with a score of 31. It attributes the gains to coding agents, vision, knowledge work and reasoning. It also says the improvement stems from advances at every stage of training, such as more efficient pretraining and a large unified reinforcement learning (RL) run.

Availability: initially the model is offered only through Tinker, with closed weights. The company frames this within the approach it described in “a safe path to open weights”. It intends to release the full weights under an Apache 2.0 license the following week, once external safety testing and red-teaming are complete. It can already be tried in the Tinker Playground and is available for fine-tuning and sampling on Tinker. Checkpoints will be downloadable once the weights are released.

General capabilities: according to the company, the model improves across the board over the first generation, especially in long-horizon autonomous work. It combines vision and agentic capabilities to incorporate visual feedback and iteratively refine websites, games and 3D designs with interactive elements. Demonstrations cited include:

• A city simulator built with Three.js, with no prefab assets. It procedurally generates buildings, textures and cars, and the user manages power, budgets and emergencies. • A PC emulator written from scratch in Python, which boots a real Linux kernel and offers a web interface to launch an instance in the browser. • A hand-drawn valley scene with origami cranes, generated in Blender and revised with improvements to the water and the chimney. • A walkthrough of a floor plan generated from a map.

Agentic coding and auto-research: the model was trained with RL on hundreds of thousands of long-horizon programming tasks. These used various agent harnesses and different context-compaction strategies. The company improved its synthetic code data pipeline and went beyond unit-test rewards. Many tasks included an agentic RL grader that judged how the model approached the problem, not just whether the code passed the tests. It also tested early checkpoints in OpenCode under the name Fledge-Alpha. According to the company, the result follows instructions better, considers changes carefully and verifies its work without being asked.

Results in the coding and auto-research table (columns: Inkling 2 Small, GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, Muse Spark Open, Kimi K3):

• FrontierCode v2.0: 25.4%, 14.2%, 28.5%, 7.6% and 17.3%. • DeepSWE v1.1: 72.1%, 63.4%, 77.0%, 60.0% and 68.5%. • SWEBench Pro v2 (Hard): 86.3% for Inkling, 82.4% for Muse Spark Open and 88.2% for Kimi K3. No data for GLM or DeepSeek. • Vibe Code Bench v1.1: 81.6%, 30.8%, 87.7%, 76.2% and 85.0%. • PostTrainBench v1.2: 31.9% for Inkling, versus 36.8% for GLM-5.3-Flash, the highest in the row. • Terminal Bench Science v0.1: 5.7%, 4.8%, 12.9%, 1.9% and 8.1%.

In these tables Inkling 2 Small comes second in several rows, but first in none of the coding rows.

Kernel optimization (PMPP-Hard): this is an external GPU kernel optimization benchmark, adapted by the company. It has 69 tasks, four independent attempts and a 10-hour budget. Each model uses its maximum reasoning setting, the PI harness and a shared compaction policy. Each solution receives two independent LLM audits, with a third review resolving disagreements. The results (average tasks solved out of 69, and pass@4) were:

• GPT-6 Astra: average of 63 and pass@4 of 63. • Kimi K3.1 Pro: 60.5 and 63. • Claude Opus 5.5: 60 and 64. • Kimi K3: 55.75 and 66. • DeepSeek V4.1 Flash: 55.25 and 62. • Inkling 2 Small: 54.25 and 64. • Mistral Large 4: 50.25 and 61. • GLM 5.3: 52.5 and 61. • GLM 5.2: 42 and 61.

The table is ordered by area under the curve (AUC) over 10 hours, with a maximum of 690 task-hours. Inkling 2 Small scores 494.27, below GPT-6 Astra (616.39), Kimi K3.1 Pro (583.86), Claude Opus 5.5 (499.65) and DeepSeek V4.1 Flash (497.36). In that table it ranks above Mistral Large 4, Kimi K3, GLM 5.3 and GLM 5.2.

Kernel auto-research (preliminary results): the task was to speed up the forward and backward training kernels of DeepSeek Sparse Attention (V3.2). It started from the public TileLang kernels, with error close to the original (strict tolerances) and no time limit. The public baseline is not tuned for B300, leaving plenty of room for improvement. The speedups apply only to these kernels, not to end-to-end training. The results were: Claude Opus 5.5, 18.80×; GPT-6 Astra, 10.26×; Inkling 2 Small, 6.02×; DeepSeek V4.1 Flash, 5.66×; GLM 5.3, 5.03×; and Mistral Large 4, 3.43×.

End-to-end auto-research case: in a task Proximal created for one of the company's internal benchmarks, Inkling 2 Small autonomously trained a Qwen3.5-4B on Tinker to forecast time series, such as weather and Wikipedia page views. It had 12 hours and a $500 Tinker budget. The model analyzed the task, wrote its own evaluator, tested datasets and recipes, and tried different SFT and RL approaches while debugging the process. After about 4.5 hours of work (the log shows 4 h 17 min) and $426 spent on Tinker, the trained model had more than doubled its forecasting score. The run log makes it possible to reconstruct the process:

• It built an evaluation with a simple baseline: carrying the latest values forward. • It converted the history up to May into 12,500 training problems. • It detected a bug in its own evaluation: it had used the same seed for all 8 forecasts, which came out identical, whereas the final grader uses 8 different seeds. • With a corrected baseline, the base Qwen3.5-4B scored 0.191. The first SFT rose to around 25%. • It noticed that evaluations were expensive ($101 spent) and cut the cost by requesting all 8 forecasts in a single call, so the prefill is done once. • It retrained on the full history up to July 31 (SFT 2, 0.269). • It tried RL with the task score as the reward, but the forecasts became overconfident and the score dropped (26% and then 23%), so it stopped the run. It confirmed that the spread of SFT 2's forecasts was already reasonable, leaving RL little to improve. • A second epoch on the same data did not help (SFT 3, 25%). • A brief additional run on only the last 12 months produced its best model (SFT 4, 0.289). Trying the last two years (SFT 5) or seven months (SFT 6) worsened the result. • It submitted SFT 4 and stopped. On unseen August and September data, the final model makes about 15% less error than the base model.

Vision: Inkling 2 Small processes images with a 292-million-parameter vision encoder based on Perception Encoder (Bolya et al., 2025). It uses 20×20-pixel patches with 2×2 spatial pooling and interleaves image patches with text tokens. The company highlights strong performance in OCR, grounding, infographic interpretation and visual reasoning. In its comparison table (against GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, Muse Spark Open and Kimi K3), Inkling 2 Small has the highest score on:

• GDP.pdf (25.2%). • MMMU Pro Standard 10 (81.6%). • PerceptionBench (64.7%). • Humanity's Sixth Sense (33.3%). • ScreenSpot Pro (83.8% original and 91.5% with Python). • BabyVision in its original version (68.6%). With Python, 93.7% for Inkling versus 92.5% for DeepSeek. • Charxiv RQ in its original version (85.6%). With Python, Kimi K3 scores 91.3%, above Inkling's 87.7%. • Chartography (36.7% original and 80.4% with Python), although with Python DeepSeek reaches 83.0%. • AA-3D (1021).

On Code Arena (WebDev) it scores 1653, below DeepSeek (1665) and Kimi K3 (1656). On Code2Video-AA it scores 1020, below Kimi K3 (1041).

DesignArena: the model uses vision in the loop to interpret and validate the visual output it generates. Pending the release of its weights, the company says it is set to be the best open-weights model on DesignArena. In the November 5 web leaderboard it appears with 1,340 points. It sits behind Muse Wonder (max, 1,366) and Muse Spark 1.3 (xhigh, 1,351), and ahead of GPT-6 Astra (max, 1,333), Kimi K3 (1,328), Claude Opus 5.5 (1,321) and Kimi K3.1 Pro (1,320). The company itself notes that the chart's bars start at 1,100, not zero, and that the comparison includes the top ten models plus three additional open-weights models.

An anecdote the company shares: it pitted the model against some of its researchers in a blind wine tasting. According to the company, the model beat them using vision alone, correctly identifying three wines (two Cabernet Sauvignons from 1999 and 2019 and a 2023 Nebbiolo) from a photo in 19 of 20 attempts. In the reproduced example, the model warns that color in a phone photo under warm light is weak evidence. It reasons by color intensity and opacity: the Nebbiolo is the lightest and most translucent, the 2019 Cabernet the darkest, and the 1999 sits in between, more muted. It also acknowledges that uneven lighting can distort the reading.

Conclusion: Inkling 2 Small is presented as an MoE model with 17B active parameters and open weights promised within days. It stands out above all in vision and posts competitive, though not leading, results in agentic coding. The company's reading of its own charts favors the model, and the kernel results are preliminary. Independent verification of these claims will have to wait for the release of the weights under the Apache 2.0 license and the results of external safety testing.