Back to Explore
NewsVibeFix editorialUpdated Nov 15, 2026

Models Are Not Bought, They Are Built: Hedra's Model Factory Takes Post-Training to the Real World

Hedra launched Model Factory on November 10: a platform that post-trains, deploys, and keeps improving custom models for robotics, manufacturing, biology, and defense. The Physics-IQ score, the 6x image-edit speedup, and the zero-retraining gripper demo are real — but all three are vendor-reported, and there is no pricing yet.

Hedra's official site graphic: the words 'General visual intelligence' beside the Hedra logo on a black and white background.

Models are not bought. They are built.

On November 10, Hedra, a research company based in San Francisco, launched a new product: Hedra Model Factory, an autonomous post-training and inference platform for building models for the real world. The press release is a single page, but the claim behind it is large — the center of gravity in model development is moving from "pre-train one general model" to "build a specialized model for a specific problem," and the process of building models is increasingly being handed over to AI itself.

Hedra describes itself briefly: an a16z-backed research company building models and infrastructure for multimodal and physical AI. Model Factory is the first time it has packaged its research capability as a service for outside teams: compute, post-training, deployment, and continuous improvement after launch, four pieces in one platform, sold to robotics, manufacturing, biology, and defense teams. Founder and CEO Michael Lingelbach summed up the ambition in one line: "Recursive self-improvement is already changing how frontier models get built. The next step is the real world."

That line deserves to be unpacked. The first half states an open secret of the last two years in frontier AI: the best models increasingly get stronger through self-improvement loops — a model generates data, evaluates itself, and trains its next generation on the results. The second half is Hedra's bet: the same loop, moved onto real-world data — video, sensors, and simulation — is worth even more, because that data is orders of magnitude larger than text, and reinforcement learning environments and simulators often need GPU acceleration just to run at scale. In other words, Hedra believes the problems general-purpose models cannot solve will be solved by "custom training on your data," and custom training is becoming something that can be partly automated.

Hedra's official site graphic: the words 'General visual intelligence' beside the Hedra logo on a black and white background.

Four stages: compute, post-training, deployment, self-improvement

Model Factory, as Hedra describes it, has four parts. The first is compute: Hedra GPU sandboxes provide flexible capacity running on its own inference engine, so teams can scale training up without locking into long-term GPU commitments. This targets a very real pain point: smaller teams cannot afford, and do not want to be locked into, long-term compute contracts, yet post-training is exactly the kind of compute-intensive work that punishes that constraint.

The second part is post-training, the core of the whole platform. Teams bring their simulation engines and data, and Hedra helps build the reinforcement learning environments and datasets they need. Hedra researchers work alongside each customer's team, using tools such as vision-language models that caption and score data or give feedback during online reinforcement learning. For companies without their own research lab, Hedra researchers also set up data capture, evaluations, and the training-to-serving loop. Note the wording here: this is not "upload your data and wait for a model." It is "researchers work with you." That tells us Model Factory is, for now, closer to research-as-a-service than to a self-serve tool.

The third part is deployment. Models run on low-latency compute, on premises or at Hedra's nearest partner data center. The release gives two examples: distilling world action models to run faster than real time without losing physical coherence, and fitting a vision-language-action model onto an edge device. Teams can also bring an existing model to serve on Hedra's inference engine. The message is that the endpoint of post-training is not a weights file, but a service that runs within latency and cost requirements.

The fourth part is self-improvement, where Hedra's whole narrative lands. Once a model is deployed, Hedra researchers analyze its production performance, retrain the model, and release only versions that pass the customer's evals. Data from each deployment flows back into the next round of training. In the release's words, researchers work alongside autonomous AI scientists that run experiments and iterate on training runs, and every deployed model feeds data back into the next version, so the factory gets faster with every model it produces. This is a flywheel promise: the longer you use it, the better the model fits your problem.

Three sets of numbers, read separately

The release offers three research results. They are the most concrete part of the announcement — and the part that most needs to be read carefully, one by one.

The first is physically consistent video. Hedra retrained its H3 model for physical consistency and scored 34.9 on Google DeepMind's Physics-IQ benchmark, compared with 30.3 for H3 Max, with comparable visual quality. A 4.6-point gain is not dramatic, but the direction matters: competition in video generation is shifting from "looks right" to "behaves right" — objects should fall when they should fall and react plausibly on collision. For robot simulation and world models, physical consistency matters more than pretty frames. The same section notes that an autonomously distilled version rendered a five-second clip in under three seconds on AMD MI355X GPUs — faster than real time.

The second is image editing. Through distillation and inference optimization, Hedra cut 2K image edits from about 22 seconds to 3.6 seconds on the same four AMD Instinct MI355X GPUs, roughly six times faster, in under a week. The details are worth noting: the baseline ran the original bf16 model at 30 sampling steps, while the faster result uses a distilled 14-step model. In other words, part of the sixfold speedup comes from taking fewer sampling steps, not from pure engineering acceleration at equal quality. That does not make the number dishonest — distillation is itself one of the core crafts of post-training — but readers should understand that the cost and the limits of the speedup depend on the evals for the specific task, not on a single multiple.

The third is the most interesting, and the easiest to misread. In a simulation study, Hedra fitted a robot running a trained policy with a new gripper and restored task success from near zero to near 100% with no retraining, by matching the new gripper's reach and grasp point in the wrist camera view. If this holds up, it matters: changing robot hardware usually means collecting new data and retraining, at high cost, so a "zero-retraining adaptation" method — even one demonstrated only in simulation — points at a key bottleneck in robot deployment. But it must be stated plainly: this is a simulation study, not a result on real hardware, and the gap between simulation and the real machine is where the hardest part of the real world still sits.

15x revenue growth, and research-as-a-service as a business

The release also contains a number that is easy to skip: Hedra grew revenue 15x over the past 12 months. There is no absolute figure and no customer name, but it explains why Model Factory exists now — Hedra wants to open up the research and infrastructure behind that growth to "teams that need custom models built on their own data." This is a classic capability spillover: build revenue with your own research first, then turn the method of doing research into a product.

The other side of this model is that it naturally serves teams that can afford a research collaboration. Robotics, manufacturing, biology, and defense — the four areas named in the release — are none of them "download and go" markets. They all have their own data, their own simulators, their own compliance requirements, and a first-hand sense that general-purpose models do not solve their problem. That is exactly what Model Factory sells: you do not need to hire your own reinforcement learning team, because Hedra's researchers, working with autonomous AI scientists, will do the job for you.

But there is a sober point here too. Hedra's team background is genuinely impressive: Lingelbach spent his career on reinforcement learning for visual and physical systems, first at Stanford and NVIDIA; the team brings together talent from Stanford, Berkeley, and UIUC, with experience at NVIDIA, Meta Superintelligence Labs, Fireworks AI, and Salesforce, and has published research in generative video and embodied AI at NeurIPS, CVPR, and ICLR. Background, however, is not delivery. Whether Model Factory can replicate the "researchers plus AI scientists" way of working across dozens of customers while honoring each customer's evals is a question the release does not answer. That is the hardest part of research-as-a-service: craft can be excellent, but craft is hard to scale.

Defense customers deserve a sentence of their own

The release lists "defense" alongside robotics, manufacturing, and biology as target customers. That is not unusual in AI infrastructure announcements, but it is still worth noting: a platform for physical AI, world models, and robot policies is dual-use by nature — the same capability that helps a model understand the physical world and helps a robot adapt to new hardware can serve a factory, a lab, or other settings. Hedra does not spell out its policy toward defense customers in the release. That is not its obligation, but for readers who care about AI governance, it is a fact to record rather than a word to skim past.

Three things the release does not say

First, price. The release mentions no pricing at all, not even a standard package. Given that Model Factory currently takes the form of "work with Hedra researchers," pricing is most likely negotiated per project. That means it is not, in the short term, something independent developers and small teams can simply use — at least not in the "sign up, swipe a card, start working" sense.

Second, independent verification. The Physics-IQ score of 34.9, the sixfold speedup, and the zero-retraining gripper recovery are all numbers reported by Hedra itself. Physics-IQ is a public benchmark from Google DeepMind, which adds credibility; but the image-editing speedup and the gripper experiment have no public third-party reproduction. Readers should treat these numbers as vendor-reported results, not as industry consensus.

Third, what "available" really means. The release says Model Factory is "available starting today" at the new hedra.com, and that teams can work with Hedra researchers, start with flexible GPU sandbox capacity, or serve an existing model on its inference engine. Those are three different thresholds: research collaboration is the heaviest, the sandbox the lightest. The middle option may be the most meaningful for most teams — rent compute first, run your own post-training, then decide whether to hand over the research collaboration too. But the release does not explain the sandbox's entry conditions or billing, and those details will have to wait for the website to say more.

Three things to watch next

Read the release, but hold the verdict. Whether Model Factory is a business or a narrative will come down to three things over the next three months. First, customer cases. The release names no customer and cites no number from a real production line. The Physics-IQ score and the gripper experiment both come from Hedra's own lab, and between lab results and production results sit all the things a press release never mentions: data distribution shifts, hardware wear, operator habits. The first customer willing to put its name on the record will say more than any new benchmark score.

Second, the sandbox threshold. Of the three ways to use Model Factory, the GPU sandbox is the lightest, and the most likely to reach smaller teams first. But the release says nothing about how the sandbox is priced, what the minimum budget is, or how long the queue runs. If those details stay unpublished, or the pricing only makes sense for large customers, the "custom models for everyone" story deserves a discount. If the sandbox turns out to be genuinely sign-up-and-go, pay-as-you-go, that would be a bigger signal than the research partnership itself, because it would mean post-training is moving from project work to product.

Third, independent reproduction. Physics-IQ is a public benchmark, so others can run it; the image-editing speedup and the gripper experiment are far harder to reproduce. If an independent team or academic group reproduces any of them in the coming months, Hedra's credibility takes a step up. If nothing appears, readers should keep treating these numbers as vendor-reported figures. None of this needs daily tracking — checking the website and the news once a month is enough. If there is follow-through, the business is moving; if there is not, it is still a roadmap.

What it means for ordinary developers

If you are an independent developer or a small team, Model Factory is probably not built for you today. But the direction it points to is worth remembering: the next stage of model competition may not be "whose general model is bigger," but "who can turn a general model into your specialized model most cheaply and quickly." Post-training, distillation, evaluation, deployment, retraining — this used to be the exclusive work of large-company research teams, and it is now being packaged as a service. Hedra is one of the earlier and more complete movers in that packaging.

For people working on robotics and embodied AI, the announcement deserves a closer read. Physical consistency, real-time distillation of world action models, and zero-retraining hardware swaps — these three points land exactly on three pain points of getting embodied AI into production: simulation must be trustworthy, inference must be fast enough, and hardware iteration must not mean starting training from scratch. Hedra's examples are not random; they read like a list of "here is what we can solve." If you are building in this space, that list is worth checking against your own roadmap; if you are only watching, it can help you judge that the bottleneck in embodied AI is shifting from "model capability" to "the engineering of getting models into production."

There is one more angle worth adding: Hedra's decision to launch Model Factory loudly, right now, is itself a market signal. Over the past year, video generation and world models have improved steeply, but the engineering that turns a model into a working system — evaluation, distillation, deployment, continuous improvement — has not kept up. Whoever productizes that engineering first may capture the next wave. Hedra is betting on exactly that gap: models will keep getting cheaper, and the ability to customize models will keep getting more valuable.

Finally, back to "recursive self-improvement." The idea is no longer new in frontier model circles, but in physical AI and robotics it is still early. Hedra's bet is that when the data is video, sensors, and simulation, the self-improvement loop still holds — and pays off even more. Whether that bet is right will depend on what comes next: not benchmark scores in a press release, but customer cases — a real team using Model Factory to build a model that runs on a real production line. Until then, Model Factory reads more like a well-written roadmap: the direction is clear, and the evidence is still on its way.

Views 6Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A scientist in a white coat and safety glasses examines a sample under a microscope in a bright laboratory, holding a tablet
News
Great at Assembling Molecules, Weak at Picking Drugs: OpenAI's MedChemBench Grades AI's Medicinal Chemistry

On 20 November 2026, OpenAI released MedChemBench: 99 medicinal-chemistry problems from 117 papers and patents, covering 761 molecules. GPT-6.2 Astra scores 84.5% overall — but 95.4% on assembling structures and only 50.4% on choosing compounds. The total flatters; the components tell the truth.

Industry TrendsModel UpdatesTesting & Quality
A developer with glasses types on a laptop late at night at a wooden desk, a large monitor behind showing colourful code, blurred city lights outside the window
News
Run the Middle Layers Three Times: Inkling 2 Mini Puts a 26B Coding Model on Your Laptop

On 20 November 2026, Thinking Machines Lab released Inkling 2 Mini: a 26.4B-parameter looped transformer that runs its middle 12 layers three times per token, fits in about 20GB, and runs offline on a laptop. It leads open-weights models of its size on DeepSWE and SWE-bench Verified — with the caveats printed in the same announcement.

Product LaunchModel UpdatesAI Coding
A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates