# Riemann 13.1B Model: NLL Falls 1.323 to 1.310 [2026] > Source: > Published: 2026-11-06 19:30:41+00:00 Riemann Computing Inc. published a technical account on October 28, 2026 that explains how it grew a 163-million-parameter checkpoint into *My Name is Riemann*, a decoder-only language model with 13.1 billion parameters. The account came through the OpenPeer AI organization on Hugging Face, and the [model card for the prerelease](https://huggingface.co/OpenPeerAI/My-Name-is-Riemann-Prerelease/blame/4b08de9a557e2b55a6b21ebb1f3bb2e160a34144/README.md) lists the exact count at 13,110,702,080 parameters. The Riemann Computing technical account matters less for the headline size than for the method. The team did not train a 13.1B network from random weights. It started from the published weights of OpenPeerAI/OpenPeerLLM, then climbed a six-rung growth ladder with a custom tool called OpenPeer-AI-NTK-Trainer. The account also discloses hardware, an evaluation set and a before-and-after result for a stage-six controller. We covered the October prerelease and its spec sheet in our earlier piece on [Riemann’s 13.1B-parameter model reaching Hugging Face](https://shattered.io/riemann-computing-13b-ai-model-prerelease-demo-2026/). This analysis skips that ground and asks a narrower question: what does the new technical account prove, what does it leave open, and what should developers and buyers of open weights do with it today. 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 Published on October 28 The technical account is a write-up of the development process, released by the OpenPeer AI organization on Hugging Face. Riemann Computing also pushed the story through press channels. A PRWeb release carried the headline [Riemann Computing Releases 13B Parameter AI Model Demo, Aims at Disrupting Model Training Costs](https://www.prweb.com/releases/riemann-computing-releases-13b-parameter-ai-model-demo-aims-at-disrupting-model-training-costs-302916102.html), and a later release on Newsfile Corp used the title Riemann Computing Publishes Technical Account of 13.1-Billion-Parameter AI Model Development. Three facts anchor the account. First, the model is a prerelease and the company labels it stage six of its growth ladder. Second, the complete loadable model sits on the Hugging Face Model Hub under the OpenPeerAI organization. Third, the work grew an existing small model instead of training a new one, which is the cost-saving claim the company keeps returning to. No named Riemann spokesperson appears on the record in the material we reviewed. The fact sheet we worked from carries no direct quotation and no price. This article therefore paraphrases documents and does not quote individuals. When a number appears below, it comes from the model card, the technical account or the press releases. When a number is our own arithmetic, we say so. ## The Spec Sheet at a Glance The table below collects the configuration the company published. Every row is a reported value, not an independent measurement. The shard average at the bottom is our division of the total size by the shard count. | Item | Reported value | |---|---| | Company | Riemann Computing Inc. | | Model | My Name is Riemann (prerelease, stage six) | | Type | Decoder-only language model | | Parameters | 13,110,702,080 (about 13.1 billion) | | Base checkpoint | OpenPeerAI/OpenPeerLLM, 163 million parameters | | Hidden size | 5,120 | | Layers and heads | 40 transformer layers, 40 attention heads | | Feed-forward size | 20,480 | | Context length | 2,048 tokens | | Precision and storage | FP32, 52.4 GB across eleven safetensors shards | | Final-stage data | 2.16 billion tokens from a 2.30-billion-token corpus | | Training dates | October 19 to 20, 2026 | | Average shard size (our math) | About 4.8 GB | The numbers cross-check cleanly. A feed-forward width of 20,480 is exactly four times the hidden size of 5,120, which is the standard ratio for a dense transformer. Four bytes per FP32 weight across 13,110,702,080 parameters gives about 52.4 GB, matching the stated shard total. Forty layers at that width also land near 12.6 billion parameters in the attention and feed-forward blocks alone, and the embedding tables plausibly account for the rest. Nothing in the spec sheet looks inflated. ## How Staged Growth Took 163 Million Parameters to 13.1 Billion Staged growth means a team trains a small model, expands its width or depth, initializes the larger network from the smaller one, trains again, and repeats. Riemann Computing calls its sequence a growth ladder and says the prerelease sits on the sixth rung. The company reports successive staged growth in place of a fresh run, and it positions the approach as a way to cut what training costs. ### The 80-fold jump in plain numbers Divide 13,110,702,080 by 163 million and you get a growth factor near 80. That is a large leap, but the ladder spreads it over six stages, so no single step multiplies the network by anything close to 80. Each rung can reuse what the previous rung learned, which is the whole point of the technique. The final stage is the one with public numbers. It used 2.16 billion tokens drawn from a 2.30-billion-token corpus, which is about 94 percent of the available text. The company reports training dates of October 19 and 20, 2026, so the whole top-rung run fits inside two days according to the account. ### Why hardware changed between rungs One detail from the technical account stands out. Early training ran on a laptop with an external GPU. Later training moved to a cloud node with eight AMD Instinct MI325X GPUs. That split makes sense for a ladder. The small rungs fit on hobbyist hardware, while the top rung needs a multi-GPU server to hold and update a 13.1B network. The hardware disclosure also fills a gap. When the prerelease first appeared, the company had not said what machines produced it. The account now ties the expensive part of the work to a specific AMD accelerator node, which matters to readers tracking how much of the open-model scene runs on non-Nvidia silicon. ## The Stage-Six Controller and OpenPeer-AI-NTK-Trainer The account credits a tool named OpenPeer-AI-NTK-Trainer for the training work. NTK refers to neural tangent kernel methods, a family of techniques that describe how wide networks behave during training. Riemann Computing uses the trainer in place of conventional adapter [fine-tuning](https://shattered.io/category/ai-machine-learning/), and its press material says the method aims to lower the cost of training and growing models. The technical account reports a before-and-after test around a stage-six controller. The company measured the model on a held-out set, then measured it again after applying the controller. That pairing is the closest thing to a controlled result in the whole release, so it deserves a close read. The next section does that. We cannot independently describe the controller’s internals beyond what the company published. The fact sheet gives us the trainer name, the controller label and the two result sets. It does not give us a paper, a hyperparameter table or a public ablation. Anyone planning to reproduce the method should treat the Hugging Face repository and the trainer’s source as the working documentation and expect gaps. ## Reading the Evaluation: NLL Falls From 1.323 to 1.310 The evaluation set holds 1,000 held-out prompt-and-completion pairs. Applying the stage-six controller cut negative log-likelihood from 1.323 to 1.310. Token accuracy rose from 67.73 percent to 67.84 percent. The table adds our own conversions so readers can see the size of each move. | Metric | Before controller | After controller | Change | |---|---|---|---| | Negative log-likelihood | 1.323 | 1.310 | Down 0.013 (about 1.0 percent) | | Token accuracy | 67.73% | 67.84% | Up 0.11 points | | Implied perplexity (our math) | About 3.755 | About 3.706 | Down about 1.3 percent | | Evaluation size | 1,000 pairs | 1,000 pairs | Same held-out set | The perplexity row comes from exponentiating the NLL values, which is a standard conversion. The snippet below reproduces every derived figure in this article so readers can check the arithmetic. ``` python import math params = 13_110_702_080 base = 163_000_000 print(round(params / base, 1)) # growth factor, 80.4 print(round(params * 4 / 1e9, 1)) # FP32 size in GB, 52.4 print(round(2.16e9 / params, 3)) # final-stage tokens per parameter, 0.165 print(round(math.exp(1.323), 3), round(math.exp(1.310), 3)) # 3.755 3.706 ``` ### What a 1,000-pair test can and cannot prove The gains are small. A 0.013 drop in NLL and a 0.11-point rise in accuracy show the controller moved the numbers in the right direction. They do not show the improvement would survive a different test set. A 1,000-pair slice is a thin sample, and the account gives no confidence interval or repeat-run spread in the material we saw. The Hugging Face model card also marks its evaluation entries as not verified in its metadata. That flag does not mean the numbers are wrong. It means no outside party has confirmed them. For a prerelease from a small lab, that is a normal state, and it tells readers to wait for third-party runs before treating 67.84 percent as a benchmark score. ## Why 2.16 Billion Training Tokens Is a Small Budget The final stage used 2.16 billion tokens on a 13.1-billion-parameter network. That works out to about 0.165 tokens per parameter. DeepMind’s Chinchilla work set the reference point many labs still cite for compute-optimal training, and it called for far more tokens per parameter than this run used. Riemann’s ratio sits well below that guidance. That gap is not a mistake. A grown model inherits knowledge from earlier rungs, so it does not need to learn everything at the top. The cumulative token count across all six stages is higher than the final stage alone, but the account’s own headline claim is that growth beats a scratch run on cost. The flip side is that the model has seen limited text at its full size, and that shows up in capability limits. A rough compute estimate puts the scale in context. The common approximation of six floating-point operations per parameter per token gives about 1.7 times ten to the twentieth power operations for 13.1 billion parameters and 2.16 billion tokens, if every weight updated at full size. That figure is our arithmetic, and it overstates the cost if the trainer updates only a subset of parameters. It still shows why a two-day run on eight accelerators is plausible. ## Memory Footprint: Who Can Actually Run a 13.1B Model Riemann ships the prerelease in FP32, which is the heaviest common format. At 52.4 GB, the weights alone overflow every consumer graphics card on sale. The table shows what the same parameter count would need at lower precision. Only the FP32 row is a published figure. The others are our weights-only arithmetic and leave out the key-value cache and activations. | Precision | Bytes per weight | Approximate weight size | Source | |---|---|---|---| | FP32 | 4 | 52.4 GB | Published by Riemann | | BF16 or FP16 | 2 | 26.2 GB | Our arithmetic | | 8-bit | 1 | 13.1 GB | Our arithmetic | | 4-bit | 0.5 | 6.6 GB | Our arithmetic | The takeaway for hobbyists is simple. Nobody should plan to run the FP32 files on a gaming card. A quantized copy would bring the model within reach of mid-range hardware, but the company has not announced one in the material we reviewed. Community conversions usually follow quickly for open weights, though we cannot promise one here. The 2,048-token context length is the second practical limit. That window suits short completions, small code snippets and experiments. It does not suit long documents or agent loops that carry large tool outputs. Teams evaluating the model for production work should test that ceiling first. ## Historical Context: Growing Models Before Riemann The idea of expanding a trained network instead of restarting is not new. Researchers have studied function-preserving expansion for years, with Net2Net often cited as an early example, and later teams applied depth and width expansion to language models. The common thread is that a larger network can start life behaving like a smaller trained one, then diverge as training continues. What makes Riemann’s account unusual is the starting point and the stage count. A 163-million-parameter base is tiny by current standards, and six rungs to 13.1B is a long ladder. Most public expansion work starts from a base that is already several billion parameters. Starting this small means the early rungs can run on a laptop, which fits the hardware story in the account. For readers who track the broader field, our [AI and machine learning coverage](https://shattered.io/ai-machine-learning/) follows the open-weight releases that set the bar this prerelease has to clear. Our ranking of the [best open-source LLMs for October 2026](https://shattered.io/best-open-source-llm-october-2026-kimi-glm-deepseek/) shows how quickly the mid-size tier moves, and a small lab entering it needs more than a size claim. ## Market Impact: Where a 13.1B Prerelease Fits The model arrives in a market where the headline numbers keep getting bigger. Tencent’s Hy4 carries 770 billion parameters under an Apache 2.0 license, as we reported in our piece on [Hy4’s 770B open-source release](https://shattered.io/tencent-hy4-770b-parameters-apache-2-0-open-source-model-2026/). By our division, that is about 59 times the size of Riemann’s model. A 13.1B network does not compete on raw capability with that class. Mixture-of-experts designs blur the size comparison. Baidu’s ERNIE 5.5 Flash Base has 280 billion total parameters but only 14 billion active, which we covered in our look at [ERNIE 5.5 Flash Base](https://shattered.io/baidu-ernie-5-5-flash-base-280b-moe-14b-active-2026/). Its active compute per token lands near Riemann’s dense 13.1B, so the two compete for similar inference budgets even though their total sizes differ by a factor of about 21. The commercial angle is the cost claim. If staged growth really cuts the bill for building a mid-size model, small labs and universities gain a path into a tier that used to need a large cluster. Riemann has not published a dollar figure, a GPU-hour count or a price, so the market cannot yet price that advantage. Buyers should treat the cost story as a hypothesis until the numbers appear. ## Competitive Comparison: Growth, Scratch Training and Experts Three build strategies now compete for the open-weight crown. Each trades compute, quality and complexity differently. The comparison below uses general properties of each approach, not benchmark claims, because no head-to-head test of Riemann against these designs exists in the material we reviewed. **Staged growth** reuses earlier work and lets a team spend most of its compute on the top rung. Riemann is the example here. The risk is that the final model inherits the weaknesses of a very small base. **Scratch training** avoids that inheritance but demands the full token budget at full size, which is why frontier labs rely on large clusters. **Mixture-of-experts** keeps active compute low while total parameters climb, as Moonshot’s 2.8-trillion-parameter model shows in our report on [Kimi K3.1 Pro going open-weight](https://shattered.io/moonshot-kimi-k3-1-pro-open-weights-2-8-trillion-parameters-2026/). That model is about 214 times the size of Riemann’s by raw count, though its active parameters are far fewer. Riemann wins on accessibility of the method, because a laptop-first ladder invites imitation. It loses on context length and on proven quality. Until outside testers run standard suites against the weights, the model competes on story and cost promise more than on measured results. ## Source Check: October 28 Versus November 4 Dates differ across coverage, and readers should know why this article uses October 28. The fact sheet places the technical account on October 28, 2026 through the OpenPeer AI organization on Hugging Face. A write-up on Emergent, titled [Riemann Computing Unveils 13.1B-Parameter AI Model Details](https://emergent.sh/news/riemann-computing-unveils-13-1b-parameter-ai), says the documentation was released on November 4 and carries a November 5 posting date. The likeliest explanation is lag. The primary publication came first, and secondary outlets caught up days later. We follow the primary date. The Emergent piece also describes the architecture and training in general terms, with no figure we could match to the model card beyond the 13.1-billion headline. We use it only to document the date gap. Two items stay unconfirmed. The company has not published a price in the records we reviewed. No named spokesperson gave an on-record quote that we could verify. Both gaps matter for a news story, and we flag them so no reader mistakes silence for a reported fact. One naming quirk also deserves a mention. The model card’s configuration labels for stage six carry a 12.85b tag, while the headline count is 13.1 billion. The exact parameter figure of 13,110,702,080 appears in the card, so the tag likely reflects an internal label from an earlier build. The company has not explained it, and we flag it as a loose end. ## Five Predictions for My Name is Riemann These are our expectations, not reported facts. We built them from the gaps in the record. 1. **Outside testers will run standard benchmarks within weeks.** Open weights on Hugging Face draw community evaluations fast, and the metadata flag on the company’s own numbers invites them. Expect scores that look lower than the company’s held-out figures, because held-out slices from a model’s own data pipeline tend to flatter it. 2. **The 2,048-token context will draw the first complaints.** It is short for 2026, and users will push the limit as soon as they load the weights. 3. **Quantized community builds will appear.** A 52.4 GB FP32 release is awkward, and open-weight models of this size usually get 4-bit and 8-bit conversions from volunteers. 4. **Other small labs will copy the ladder.** The laptop-first start lowers the barrier, and any lab with one multi-GPU node can try a similar climb. 5. **A later rung will follow.** The prerelease label signals unfinished work, so expect either more training at the current size or another growth step. ## What Developers and Buyers Should Do Now Treat the release as a research artifact and not a production dependency. The checklist below keeps expectations in line with the evidence. - Pull the model card and read the configuration, including the stage labels, before downloading 52.4 GB. - Run your own evaluation on prompts from your domain, since the company’s 1,000-pair set tells you little about your workload. - Test the 2,048-token window against your longest realistic input. - Check the license terms on the repository before any commercial use, because the fact sheet does not confirm them. - Wait for independent benchmark results before comparing the model against the mid-size open tier. - Ask Riemann for GPU-hour and cost figures if the cost claim drives your interest. The takeaway: the technical account turns a vague cost pitch into a documented process with a named trainer, named hardware and a small but real evaluation result. It does not yet prove the process beats other ways of building a 13B model. That proof needs outside runs and published cost data. ## Frequently Asked Questions **What did Riemann Computing publish on October 28, 2026?** It published a technical account of how it developed My Name is Riemann, a 13.1-billion-parameter decoder-only model. The account came through the OpenPeer AI organization on Hugging Face. It covers the staged growth method, the training tool, the hardware and an evaluation of a stage-six controller. **How many parameters does My Name is Riemann have?** The model has 13,110,702,080 parameters, which the company describes as 13.1 billion. It grew from the 163-million-parameter OpenPeerAI/OpenPeerLLM checkpoint, a jump of roughly 80 times by our calculation. **What is staged growth in language models?** Staged growth trains a small model, expands it, initializes the larger network from the smaller one and trains again. Riemann Computing used six stages and calls the sequence a growth ladder. The prerelease is stage six. **How good is the model?** The company reports negative log-likelihood of 1.310 and token accuracy of 67.84 percent on 1,000 held-out pairs after applying the stage-six controller. Those figures are self-reported, and the model card marks them as unverified. No independent benchmark result was available in the material we reviewed. **What hardware trained it?** Early stages ran on a laptop with an external GPU. Later training used a cloud node with eight AMD Instinct MI325X GPUs. **Can I run it on a home PC?** Not in its released form. The FP32 weights take 52.4 GB across eleven shards. A quantized copy could fit smaller hardware, but the company has not announced one, and the 2,048-token context limits long-document use. **How much does it cost?** No price appears in the records we reviewed. The weights are listed on the Hugging Face Model Hub, and readers should check the repository for license and access terms before use.