Thinking Machines released the open weights for Inkling 2 Small on November 12, 2026, making its 594-billion-parameter model downloadable under Apache 2.0. Developers can now obtain the checkpoints for research, fine-tuning and third-party integration, alongside existing platform access and new deployment options. The practical catch is substantial: the company lists minimum aggregate GPU memory requirements of 500GB for NVFP4 and 800GB for MXFP8. (Announcement, model card)
The date matters. Inkling 2 Small originally launched on November 5 with closed weights through Thinking Machines’ Tinker platform. The November 12 update is the downloadable-weights release, opening a different set of choices for developers who need control over model hosting or customization.
What Changed on November 12
Thinking Machines added a dated update to its original announcement stating: “Inkling 2 Small’s weights are now available for download on HuggingFace.” The same update lists inference integrations, post-training tools and seven third-party API providers. (Thinking Machines announcement)
The original November 5 post had promised an Apache 2.0 release the following week after external safety testing and red-teaming. The November 12 update documents the download becoming available; it should not be read as a detailed report of those evaluations.
The model card identifies Apache 2.0 as the license and explicitly describes research, fine-tuning and integration into third-party products as purposes of the open-weights release. That gives developers a deployment path beyond calling a company-operated endpoint. (Thinking Machines model card)
My view is that control is the important change here. A hosted API gives an application access to outputs. Downloadable weights let a team choose its serving stack, manage deployment within its own infrastructure and work on model customization. Those options matter even when the hardware needed to exercise them remains expensive.
What Inkling 2 Small Supports
Inkling 2 Small is a multimodal autoregressive transformer with a sparse Mixture-of-Experts, or MoE, architecture. Its main published specifications are:
- 594 billion total parameters, with 17 billion active parameters.
- A 48-layer decoder-only transformer, routing each token to 8 of 384 experts.
- A hybrid of local and global attention layers.
- Text and image inputs, with text output.
- A context window of up to 1 million tokens.
- Image dimensions between 40 and 4,096 pixels recommended for optimal performance.
These specifications come from the model card. Multimodal support here means the model can process images alongside text. It is not an image-generation model, and its listed output modality is text.
Why 17B Active Still Requires a Large Cluster
MoE routing selects a subset of experts for each token, reducing the amount of parameter computation needed for that token. It does not turn the downloadable model into a 17B-parameter checkpoint: different tokens can need different experts across the much larger model.
That distinction explains why the active-parameter figure and the memory requirement can look so different. Active parameters help describe the computational work; the complete set of weights still drives a substantial storage and deployment burden.
The million-token context is another capacity to plan around. Long inputs require processing and runtime memory in addition to loading the weights. I would treat the published GPU-memory figures as deployment entry requirements, then size for the intended context length, concurrency and serving framework. The model card does not supply a configuration-specific guarantee for every combination of those settings.
NVFP4 Versus MXFP8: A 60% Memory Difference
The main Hugging Face repository supplies the NVFP4 checkpoint. Thinking Machines also links an MXFP8 checkpoint from its model card.
| Checkpoint | Published minimum aggregate VRAM | Example configurations listed by Thinking Machines |
|---|---|---|
| NVFP4 | 500GB | 2× NVIDIA B300 or 4× NVIDIA H200 |
| MXFP8 | 800GB | 4× NVIDIA B300 or 8× NVIDIA H200 |
These are company-provided hardware requirements and example configurations, not independently tested deployment results. Aggregate VRAM means GPU memory across the cluster, not ordinary system RAM.
NVFP4 is particularly important because Thinking Machines says the model was pretrained, midtrained and post-trained in that format. It is the model’s native training precision, and the company recommends that checkpoint for all use cases. (Thinking Machines model card)
The published figures reveal a useful deployment tradeoff: MXFP8 raises the stated minimum memory requirement by 300GB, or 60%. Thinking Machines reports a marginal evaluation improvement of 1–2 percentage points for MXFP8, while saying it does not consider that improvement meaningful for practical use. That performance assessment is the vendor’s, not an independent measurement.
I’d start with NVFP4 unless an application-specific evaluation gives a clear reason to choose MXFP8. Native training precision, a lower published memory threshold and the vendor’s own recommendation all support that choice. A small reported evaluation difference is a weak reason on its own to provision a larger cluster.
The 60% figure describes the difference in stated VRAM requirements. It does not imply that MXFP8 costs exactly 60% more to operate: GPU configurations, rental rates, throughput and utilization all affect the actual bill.
Serving, Fine-Tuning and Hosted Access
Thinking Machines lists Unsloth, SGLang, vLLM and TokenSpeed as integrations in the November 12 update. Its model card also lists Hugging Face among the software options for running the weights, alongside the relevant framework dependencies. (Announcement, deployment requirements)
Post-training support is listed separately through Miles, SkyRL and Snowflake Cortex Training. Tinker remains available for fine-tuning and sampling, while Tinker Playground provides inference-only access.
Separating these categories helps avoid a common setup mistake: a serving integration and a post-training platform solve different problems. The presence of both gives developers more options, but it is not a promise that one installation supports every workflow.
For teams without a suitable GPU cluster, Thinking Machines lists seven third-party hosted API paths:
- NVIDIA
- Fireworks
- Together
- Baseten
- Nebius
- Modal
- Cloudflare
That availability is reported by Thinking Machines. The model card directs developers to each provider’s documentation, so the list should not be treated as proof of identical endpoint features, context limits or quotas.
Frequently Asked Questions
3 questions
1When Did Inkling 2 Small’s Weights Become Available?
Thinking Machines announced downloadable weights on November 12, 2026. The original model launch was November 5.