# Riemann’s 13.1B-Parameter AI Model Hits Hugging Face [2026] > Source: > Published: 2026-10-23 22:45:14+00:00 Riemann Computing Inc. published a prerelease of a 13B parameter AI model on October 21, 2026, and says the training method behind it points toward lower model training costs. The model, named **My Name Is Riemann**, carries 13,110,702,080 parameters and sits on the Hugging Face Model Hub under the OpenPeerAI organization. A company press release distributed through EIN Presswire and carried by the Akron Beacon Journal’s press-release section put the announcement into circulation that evening. The details that are checkable look solid. The architecture, file sizes, token counts and training dates all appear in the public repository. The pitch is another matter. The company has not published a dollar figure, a GPU-hour count or a benchmark table, so the headline claim about cheaper training stays an assertion for now. This analysis separates what Riemann Computing has documented from what it has only promised. The release lands in a crowded month for open weights. Our [ranking of five open-source LLMs from $0.15 input](https://shattered.io/best-open-source-llm-october-2026-kimi-glm-deepseek/) shows how fast the mid-size field moves, and a prerelease from a small lab has to be read against that backdrop. Below, we walk through the spec sheet, the training record, the cost argument, the competitive picture and five predictions. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. [Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F) ## What Riemann Computing Announced on October 21 The announcement is short on drama and long on specifics about the artifact. Riemann Computing Inc. says it created a prerelease version of a 13B model and demonstrated it publicly. The weights live on the Hugging Face Model Hub under OpenPeerAI, the organization the company uses for its OpenPeer AI family of models. A basic demo, titled My Name is Riemann Prerelease basic demo, runs the Stage 6 checkpoint so anyone can test it without downloading 52.4 GB of files. The company labels the model a prerelease, and that word matters. It is not a finished production model, and Riemann Computing does not market it as one. The release invites inspection, experimentation and feedback. That framing is honest, and it also lowers the bar the model has to clear on day one. The press release is the clearest news-style coverage we could locate. It carries an October 21, 2026 dateline from San Francisco and frames the project around one idea: using an NTK-based training technique instead of LoRA to cut what it costs to train and grow language models. The Hugging Face repository, the OpenPeerAI GitHub organization and the press-release syndication are primary or near-primary sources. None of them is an independent technical evaluation, and readers should weigh them accordingly. ## My Name Is Riemann Spec Sheet: 13.1B Parameters at a Glance Start with the numbers the company published. They describe a conventional dense, decoder-only transformer, which means every listed parameter participates in processing each token. Nothing in the retrieved material describes a mixture-of-experts design. The table below collects the configuration as reported by Riemann Computing and its Hugging Face repository. | Specification | Reported value | |---|---| | Developer | Riemann Computing Inc. (formerly Stark Drones) | | Model name | My Name Is Riemann (prerelease) | | Exact parameter count | 13,110,702,080 | | Architecture | Decoder-only language model | | Hidden size | 5,120 | | Transformer layers | 40 | | Attention heads | 40 | | Feed-forward network size | 20,480 | | Context window | 2,048 tokens | | Weight precision | FP32 | | Files | 11 Safetensors shards, 52.4 GB total | | Training dates | October 19 and 20, 2026 | A quick sanity check shows the figures hang together. At four bytes per FP32 weight, 13,110,702,080 parameters come to roughly 52.4 GB, which matches the shard total the repository lists. The feed-forward size of 20,480 is exactly four times the hidden size of 5,120, a ratio that is standard for dense transformers. These are small consistency checks, but they tell us the published numbers describe a real checkpoint and not a marketing placeholder. One figure stands out as modest by 2026 standards. A 2,048-token context window is short next to what many current open models offer, so My Name Is Riemann suits short completions and experiments better than long-document work. ## How the Model Was Built: Six Stages and 6.5 Billion Tokens Riemann Computing describes development as a staged process. Cumulative training came to 6.5 billion tokens across six development stages, and the demo serves the Stage 6 checkpoint. The company trained the prerelease on October 19 and 20, 2026, so the whole effort sits inside a two-day window as described in the release. ### The Final-Stage Numbers The last stage drew 2.16 billion tokens from a 2.30-billion-token corpus. That is about 94 percent of the available text, a high draw rate that suggests the team used most of the corpus rather than sampling a thin slice. If that stage fit inside the 48-hour window, the average throughput works out to at least 12,500 tokens per second. The company has not disclosed its hardware, so we cannot say whether that rate is impressive or ordinary. ### Progressive Growth From OpenPeerLLM According to the Hugging Face model card, the model was grown in rungs from the published weights of OpenPeerLLM, an earlier model in the same family. That detail changes how to read the 6.5 billion token total. The team did not run a fresh 13B pretraining campaign from random initialization. It expanded an existing model step by step and trained at each rung, which is a different and usually cheaper path. It also explains why a 6.5 billion token budget is plausible for a model this size. By our arithmetic, that is about half a token per parameter, far below what a from-scratch run of this scale normally consumes. The comparison only makes sense if the growth method carries over knowledge from the smaller parent. The company has not published the growth schedule, the corpus composition or the optimizer settings, so outsiders cannot yet test that assumption. ## What a 13B Parameter AI Model Needs to Run Weight format drives who can actually use this release. The full checkpoint ships in FP32, which doubles the footprint of a bf16 copy. Simple math puts a bf16 version near 26.2 GB, roughly half the 52.4 GB of the published shards. The demo already makes that trade: it loads the Stage 6 checkpoint in bf16 and runs on a ZeroGPU slot. For a developer, the practical read is simple. Casual testing belongs in the hosted demo. Local experiments need a machine with enough memory to hold tens of gigabytes of weights, plus overhead for activations. Our look at [Nvidia versus AMD GPUs for local AI](https://shattered.io/nvidia-vs-amd-gpu-local-ai-32gb-vs-24gb-2026/) covers how memory size shapes which models fit on consumer cards, and the same logic applies here. Quantizing the weights would shrink the footprint further, though Riemann Computing has not announced quantized builds. The repository also ships custom code, an NTK controller and training-pipeline material, per the model card. Custom code means loading the model may require trusting remote code, a standard caution for any Hugging Face repository. Teams should read that code before running it on production hardware. ## The Basic Demo: What Stage 6 Shows and What It Does Not The demo record names three parties: OpenPeer AI, Riemann Computing Inc. and Andrew Magdy Kamal. It lists two licenses, Open Innovation License v2 and MIT. According to the model card, the Space supports article completion, a token-level surprisal view and a fluency test. Two linked demos round out the set, one comparing OpenPeerLLM and one showing the NTK trainer. Token-level surprisal is a useful inclusion. It shows how unexpected each token looks to the model, which gives testers a window into where the model is confident and where it flounders. That is more informative than a single fluent paragraph, and it suggests the team wants scrutiny, not applause. What the demo cannot do is prove the cost claim. A working checkpoint tells us the training method produces a model that generates text. It says nothing about how much compute the method saved relative to an alternative, because a demo has no baseline beside it. To judge the savings, outsiders need paired runs with logged hardware, time and loss curves. ## The Training Cost Claim: Confirmed Versus Unconfirmed The headline promise concerns model training costs. The press release frames NTK-based training as a way to cut those costs compared with LoRA, and the story title says the company aims at disrupting that spending. No retrieved record established a specific savings figure, a dollar price or an independently verified performance advantage. The table sorts the claims by how well the evidence supports them. | Claim or detail | Status | Basis | |---|---|---| | 13,110,702,080 parameters, decoder-only | Documented | Hugging Face repository and press release | | Hidden size 5,120, 40 layers, 40 heads, FFN 20,480 | Documented | Repository configuration | | 6.5 billion tokens across six stages | Company-reported | Press release and model card | | 2.16 billion of 2.30 billion tokens in final stage | Company-reported | Press release and model card | | Public demo of Stage 6 checkpoint | Documented | Hugging Face Space | | Lower training cost than LoRA | Unconfirmed | No cost figure published | | Benchmark results | Unconfirmed | None published | | Production readiness | Not claimed | Company calls it a prerelease | Missing evidence includes an audited GPU-hour total, a hardware configuration, energy use, a cloud bill and a cost-per-token comparison. Without those, nobody can compute a savings percentage, and any outlet repeating a number should be asked where it came from. Our own fact check turned up none. ## NTK Training Versus LoRA: What the Company Is Pitching LoRA is the widely used adapter approach. It freezes the base weights and trains a small set of added low-rank matrices, which keeps memory and compute needs low for [fine-tuning](https://shattered.io/category/ai-machine-learning/). Riemann Computing says it used its own NTK gate controller instead of LoRA to grow and adapt its model. The retrieved material does not unpack the acronym. In machine learning literature, NTK usually stands for neural tangent kernel, but the company has not spelled out its meaning here, so treat that reading as an assumption. The pitch is that a gate controller can steer which parts of a growing network learn, so each rung costs less than a conventional full pass. That is a testable idea. A neutral lab could grow the same parent model with both methods under matched hardware and token budgets, then compare final loss and downstream scores. Until someone does, the comparison to LoRA remains a position and not a result. There is a subtler point. LoRA and progressive growth solve different problems. LoRA adapts a model that already exists, while growth makes a model bigger. A fair test needs to match the goal, not just the price tag. Readers should watch for whether Riemann Computing publishes a like-for-like experiment or keeps the comparison at the level of messaging. ## Who Is Riemann Computing? From Stark Drones to OpenPeer AI Riemann Computing Inc. was formerly known as Stark Drones. The Hugging Face organization page describes a startup centered on building technologies aimed at disrupting telecom and infrastructure as industries, and says the company stands behind the OpenPeer AI family of models. The company’s investor-facing site describes a focus on compressing network packets and building proprietary compute modules, with a stated target of cutting data storage and processing costs by 90 percent. That target is a company goal, not a measured result. A separate EIN Presswire release from July 2026 described a new programming language for physicists, and noted that the company’s core focus is data compression with AI techniques layered on top. Read together, these items paint a small company with a wide set of ambitions. The language model is one product among several, and the cost-cutting theme links them all: compress the data, shrink the compute, trim the bill. That context shapes expectations. A lab whose main business is telecom and compression is not racing OpenAI or Google for the frontier. It is using a language model to show off a training technique. Judge the release by whether the technique holds up, not by whether the model tops a leaderboard. ## Competitive Comparison: Where 13.1B Sits Among Open Models Scale alone does not rank models, but it frames the conversation. My Name Is Riemann is a dense 13.1B checkpoint. Around it, October 2026 brought much larger and more specialized open releases. The table compares headline size figures drawn from our earlier coverage and a Prompt20 guide to Chinese AI labs. It does not compare quality, since Riemann Computing has published no benchmarks. | Model | Reported size | Design | Source of figure | |---|---|---|---| | My Name Is Riemann (prerelease) | 13.1B parameters | Dense decoder-only | Hugging Face repository | | Nvidia Nemotron 3.5 | 120B parameters | Open model family | Our Nemotron 3.5 coverage | | Recursive Flint | 40B total, 2.5B active | Mixture of experts, coding focus | Our Flint coverage | | DeepSeek V4.1-Flash (Sept. 10, 2026) | 552B total, 8B active in prefill and 16B in decode | Mixture of experts | Prompt20 guide | The contrast is useful. [Nvidia’s Nemotron 3.5 brought a 120B model to PCs](https://shattered.io/nvidia-nemotron-3-5-120b-open-models-2026/), and [Recursive Flint opened a 40B coding model with 2.5B active parameters](https://shattered.io/recursive-flint-40b-moe-coding-model-2026/). Both use their size or sparsity to chase efficiency at inference time. Riemann Computing aims at a different stage of the pipeline, the cost of getting a model trained in the first place. A dense 13.1B model activates all of its weights on every token, so at inference it costs more per token than a sparse model with a small active set. If Riemann’s method works, the win lands in training, not serving. ## Historical Context: Why Training Costs Became the Fight Training spend has become the central economic question in AI. The largest labs buy compute at a scale that few companies can match, and 2026 has made the squeeze visible. Our report on [Pichai saying compute stays tight into 2027](https://shattered.io/pichai-compute-shortage-capacity-constrained-2027-2026/) captured the mood: even the biggest buyers say demand outruns capacity. When supply is short, any method that promises fewer GPU hours draws attention. Frontier labs attack the problem with money and engineering. Our piece on [OpenAI training frontier AI offline](https://shattered.io/openai-frontier-training-no-internet-offline-world-2026/) showed one extreme, a large organization rebuilding its pipeline around control and security. Small labs attack it differently. They reuse published weights, grow models in stages and look for cheaper update rules. Riemann Computing sits firmly in that second camp. The pattern has precedent in the open-weights era. Each time a strong base model is published, a wave of derivative releases follows, built by teams that cannot afford a from-scratch run. My Name Is Riemann is one such derivative, grown from OpenPeerLLM. Whether it adds anything beyond size depends on evidence that has not yet appeared. ## Market Impact: Small Release, Large Question Direct market impact is minimal today. Riemann Computing is a private startup, the model is a prerelease, and no pricing, hosting plan or enterprise offering accompanied the announcement. No benchmark has shown it competing with the open models above. Anyone expecting a shift in cloud GPU demand from this release alone will be disappointed. The indirect impact is about narrative. Investors and buyers are watching for signs that AI progress will not depend only on ever larger clusters. A credible, reproducible method for cutting training cost would matter to every lab and enterprise that fine-tunes models, from startups to banks. The claim is attractive precisely because the market wants it to be true, which is also why it deserves a hard look. Storage and data movement belong in the same discussion. Training runs consume far more than GPU time, and our analysis of AI’s hidden storage bill, where one clip leaves 130x the data shows how the cost base stretches across the stack. Riemann Computing’s compression roots suggest it understands that angle, though the language model release does not address it directly. ## Licensing and Open-Source Questions The demo record lists Open Innovation License v2 and MIT. MIT is permissive and familiar. Open Innovation License v2 is less common, and its terms deserve a read before any commercial use. The search evidence we reviewed did not expose a clear license statement for the model weights themselves, so the safest course is to open the repository’s license field and accompanying files and read them directly. The lineage adds another layer. The model card names OpenPeerLLM as the base model, so downstream users inherit whatever terms govern that parent. A team planning to build a product on My Name Is Riemann should trace the license chain back one level and confirm the terms match its intended use. Open weights are not the same as open training. The repository ships weights, custom code, an NTK controller and pipeline material, which is more than many releases include. Still, full reproduction needs the corpus, the schedule and the hardware record, and the company has not published all of those. ## What Developers and Buyers Should Do Now Treat this as a research artifact. Run the basic demo, probe the surprisal view and see how the model handles your own text. Do not plan production workloads around it. The model has a 2,048-token window, no published benchmarks and a prerelease label from its own maker. Teams that care about cheaper adaptation have better-documented options today. Our walkthrough of a self-hosted AI coding assistant in 12 steps shows what a practical open-model deployment looks like, and the broader [AI and machine learning section](https://shattered.io/ai-machine-learning/) tracks new releases as they land. Compare any new claim against those baselines before committing time or budget. - Read the repository license field and the parent model’s terms before any commercial use. - Review the custom code before loading it, since the repository ships its own modeling and training files. - Ask Riemann Computing for GPU hours, hardware and loss curves behind the cost claim. - Run your own evaluation instead of waiting for a leaderboard entry. ## Five Predictions for Riemann Computing and Cheaper Training These forecasts are our analysis, not statements from the company or any outside expert. 1. **Benchmarks arrive within weeks.** Prerelease models that want attention usually publish scores soon after launch. If none appear by mid-November, read that as a signal about the results. 2. **Community testers will question the savings.** Researchers who run NTK-style growth against LoRA under matched budgets will post findings, and the first independent numbers will likely be narrower than the marketing language. 3. **A quantized build will follow.** The 52.4 GB FP32 footprint limits adoption, so a smaller-precision release is the obvious next step for anyone hoping to reach hobbyist hardware. 4. **The context window will grow.** A 2,048-token limit holds the model back, and the next rung in the growth ladder is the natural place to raise it. 5. **The story stays niche.** Without a published cost figure or a headline benchmark, major outlets will keep treating this as a small-lab curiosity. The takeaway: Riemann Computing has shipped a real, inspectable 13B checkpoint and a working demo. It has not yet shown that its training method saves money. That second step is the one the industry cares about, and it will take paired experiments, not press releases, to settle it. ## FAQ: Riemann Computing and the 13B Model ### What is My Name Is Riemann? It is a prerelease decoder-only language model from Riemann Computing Inc. with 13,110,702,080 parameters, published on the Hugging Face Model Hub under the OpenPeerAI organization. ### Who is Riemann Computing? Riemann Computing Inc. is a startup formerly known as Stark Drones. It builds the OpenPeer AI family of models and describes itself as focused on telecom, infrastructure and data compression. ### How much cheaper is its training method? Unknown. The company says NTK-based training lowers cost compared with LoRA, but it has not published a dollar figure, a GPU-hour count or an independent comparison. ### How big are the model files? The weights ship in FP32 across 11 Safetensors shards totaling 52.4 GB. The demo loads the Stage 6 checkpoint in bf16 on a ZeroGPU slot. ### How much data trained it? Cumulative training reached 6.5 billion tokens across six stages. The final stage used 2.16 billion tokens from a 2.30-billion-token corpus, and training ran on October 19 and 20, 2026. ### Is it ready for production use? No. The company calls it a prerelease, the context window is 2,048 tokens and no benchmark results have been published. ### What license applies? The demo record lists Open Innovation License v2 and MIT. Check the model repository’s license field and the parent model’s terms directly before any commercial use.