Hedra Model Factory

Hedra Model Factory is the platform Hedra launched on 10 November 2026. Officially it is an autonomous post-training and inference service “for building models for the real world,” aimed at robotics, manufacturing, biology and defense teams. Hedra is an a16z-backed research company in San Francisco; its founder and CEO is Michael Lingelbach. The official account @hedra_labs posted at 2026-11-10T17:30:47Z. A GlobeNewswire release dated 10:00 ET the same day describes the product and says it is available at hedra.com.

In the official materials, general-purpose models often do not know a particular robot, factory line or lab. Model Factory takes the recursive self-improvement loop that has mostly lived inside a few frontier labs and offers it to teams that bring their own simulators and data. The release splits the loop into four stages: compute (GPU sandboxes on Hedra’s inference engine; the company stresses there is no long-term GPU commitment), post-training (customers bring simulation engines and data; Hedra researchers work with autonomous agents to build RL environments and datasets, and can also help with data capture, evals and the training-to-serving loop when a customer has no research team), deploy (low-latency serving on premises or at a nearby partner data center — examples include distilling world action models to run faster than real time, or fitting a vision-language-action model on an edge device; existing models can also be served on Hedra’s engine), and self-improve (after deployment, models are retrained from production performance and only versions that pass the customer’s evals are released).

The release also cites company-reported research numbers: a physically retrained H3 scored 34.9 on Google DeepMind’s Physics-IQ, versus 30.3 for H3 Max, with comparable visual quality claimed; an autonomously distilled version rendered a five-second clip in under three seconds on AMD MI355X GPUs. On image editing, Hedra said it cut 2K edits from about 22 seconds (original bf16, 30 steps) to 3.6 seconds (a distilled 14-step model) on the same four MI355X GPUs in under a week, roughly 6×. Hedra also said revenue grew 15× over the past 12 months. Those figures come from the company; we have not independently reproduced them.

Key links:

• Product: https://hedra.com/ ;
• Official post (2026-11-10T17:30:47Z): https://x.com/hedra_labs/status/2120207555263905973 ;
• Press release: https://www.globenewswire.com/news-release/2026/11/10/3399821/0/en/hedra-launches-model-factory-an-autonomous-post-training-and-inference-platform-for-real-world-ai.html ;
• This is a new platform name from Hedra, not another name for its existing video models such as Character-3 or H3.

Why it matters

Physical-AI data is video, sensors and simulation — larger and more multimodal than text — and RL environments often need GPUs just to run at useful scale. Hedra is packaging “researchers plus autonomous agents post-train a model, then keep improving it from production data” as a buyable platform, and it published a few of its own optimization results as evidence. For robotics or factory teams that lack an in-house research lab but need models adapted to their own data, that is a new product-category entry point.

Keep the names straight: Hedra has been known for AI video (Character-3, H3 and related models). Model Factory, announced on 10 November 2026, is a post-training and inference platform, not another generative-model name. The Physics-IQ scores, 6× speedup and 15× revenue figure are company statements in the release; cite the source, and check the release for the exact setup (Physics-IQ is compared with H3 Max; the image-editing comparison is on the same hardware with stated step counts and precision).

Limits

Public materials do not give pricing, an SLA, or a list of supported base models, and there is no third-party evaluation. We have not reproduced the benchmark or latency numbers, and the release does not report trial counts or variance. Customers bring their own simulators and data, so results depend on data quality and the customer’s evals. Treat “available today” as the claim on hedra.com and in the release.

1. Read the official post https://x.com/hedra_labs/status/2120207555263905973 and the release https://www.globenewswire.com/news-release/2026/11/10/3399821/0/en/hedra-launches-model-factory-an-autonomous-post-training-and-inference-platform-for-real-world-ai.html ;
2. Use https://hedra.com/ for the current product entry point, how to apply, and deployment options (GPU sandboxes, on-prem or partner-data-center serving);
3. When citing Physics-IQ or image-edit latency, say these are company results from Hedra’s release, and check the comparison (H3 Max, step count, hardware);
4. Do not treat Hedra Model Factory as a new video-generation model name, and do not swap it with H3 or Character-3; figures on this page only restate the official release and announcement post — re-check the upstream source before citing them.