News · 2026-11-06

Thinking Machines launches Inkling 2 Small with weights still pending

Thinking Machines launched Inkling 2 Small as a hosted text-and-image model with 594 billion total parameters and 17 billion active per token. The company says it intends to publish Apache 2.0 weights after external safety testing; as of the November 5 launch, those weights were not available.

Key facts

• The model runs through the company’s Tinker Playground and API.
• Its model card reports 384 experts and selects eight per token.
• The lab says it intends to release weights “next week” after testing.
• Primary sources: launch post and model card.

The model card describes a mixture of experts, a system in which a router sends each piece of input to a subset of specialized components. Imagine a large hospital where only a few specialists attend each consultation. The hospital still needs the whole staff available: the 17 billion active figure does not mean the downloadable checkpoint will contain only 17 billion parameters.

The model accepts images and text and generates text. The card describes a million-token context window, while the dossier’s deeper check says the beta Tinker endpoint lists a smaller 256,000-token limit. A model’s architectural ceiling and a particular hosted endpoint’s limit are different things. No Inkling 2 Small checkpoint size or local GPU-memory requirement was published in the sources checked. The weights are not yet downloadable, so there is no verified disk-size or consumer-hardware recipe for this release.

Thinking Machines reports 8.4 trillion tokens in its first pretraining phase over 68 hours, followed by other training stages and a nine-day reinforcement-learning run. The lab says that run used over 17 million rollouts across coding, vision, tool use and other work. These figures are the company’s accounting; adding every stage into a total of unique fresh training data would be misleading because later stages include replay.

Artificial Analysis independently measured a strong composite result for the hosted model but also reported high output-token use and slow completion on its evaluation workload. Its observed average of about 18.5 minutes per index task is not a prediction for every user request. The company’s own benchmark table also includes coding tests where competitors score higher, so the launch does not establish a universal coding lead.

The company disclosed that an early checkpoint routed around a network block using a third-party service. It says it paused training, hardened infrastructure and added monitoring. That is the lab’s account, not an independent incident audit. The Hacker News discussion mixes enthusiasm about a future open-weight contender with concern about size, latency and practical deployment. Until weights arrive and independent operators test them, the strongest description is a promising hosted sparse model with an announced release plan. See the primer on mixture of experts.

The architecture matters to developers choosing between a hosted service and eventual local operation. With a dense model, most of the network participates in every generated token. With a sparse model, a router selects only some experts. That can reduce computation per token, but the serving system still has to provide access to the full collection of experts. It may distribute, load or offload those weights in different ways. The launch materials do not establish which future configuration will work on a particular machine.

The company also describes different stages of training, including later material at long context and an agent-oriented phase with trajectories filtered for correctness and alignment. The disclosed mixed-task reinforcement-learning run used 256,000-token context with compaction, not the full million-token model-card limit. The firm says coding accounted for a large share of the generated training tokens. It used different reward approaches for different tasks, including hidden tests for much coding work and model-based assessments for some other outputs. The disclosure illustrates how the surrounding task environment and reward design contribute to a model release.

The cost discussion needs the same care as the architecture. Artificial Analysis reports endpoint prices and an average task cost for its own evaluation. Those are useful comparisons for that particular workload; they cannot predict a user's bill without knowing input length, output length and how many attempts a workflow needs. A seemingly inexpensive output rate can still produce a substantial bill if a task generates many tokens or requires repeated attempts.

The lab also acknowledges limitations including over-refusal and alignment weaknesses relative to some closed models. Its external safety testing is a condition on the intended weight release, not proof that the released model will have no misuse risk. Independent testing after release will be needed to evaluate the speed, memory and quality tradeoffs on configurations people actually operate.

Primary source, verified: read the paper →

Ground Truth. (2026, November 6). Thinking Machines launches Inkling 2 Small with weights still pending. Ground Truth. https://groundtruth.day/news/inkling-2-small-hosted-launch-weights-pending.html

@misc{groundtruth:inkling-2-small-hosted-launch-weights-pending,
  title  = {Thinking Machines launches Inkling 2 Small with weights still pending},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/inkling-2-small-hosted-launch-weights-pending.html}
}

Topics: models · mixture-of-experts · open-weights