---
type: "post"
id: "hyperstition-open-sources-hyperkernel-megakernels-for-nvidia-b300-1028a1"
slug: "hyperstition-open-sources-hyperkernel-megakernels-for-nvidia-b300-1028a1"
title: "Hyperstition open-sources Hyperkernel megakernels for NVIDIA B300"
description: "Hyperstition has published an MIT-licensed alpha of Hyperkernel, a set of megakernels and command-line serving paths for NVIDIA B300 GPUs."
retrieval_nugget: "Hyperstition has published an MIT-licensed alpha of Hyperkernel, a set of megakernels and command-line serving paths for NVIDIA B300 GPUs. It spans LLM inference, driving, DNA generation, video world models, robotics, and protein folding, giving equipped teams public code to inspect and test rather than a turnkey service or demonstrated performance advantage."
published_at: "2026-11-06"
updated_at: "2026-11-06"
record_date: "2026-11-06"
date_kind: "published_at"
topics: ["inference","compute","open-source"]
entities: ["Hyperstition","Hyperkernel","Hyperserve","NVIDIA B300","DeepSeek-V4.1-Flash","Evo 2 7B","Matrix-Game 3.0","ESMFold"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "Hyperstition’s Hyperkernel release targets NVIDIA B300 megakernels, distinct from OpenBMB’s ForgeKernel GPU-kernel optimization agent."
prior_refs: ["https://newruntime.com/posts/openbmb-open-sources-forgekernel-a-self-evolving-gpu-6db18d/"]
source_url: "https://github.com/Hyperstitioncc/Hyperkernel"
source_urls: ["https://github.com/Hyperstitioncc/Hyperkernel"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:hyperstition-open-sources-hyperkernel-megakernels-for-nvidia-b300-1028a1"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/hyperstition-open-sources-hyperkernel-megakernels-for-nvidia-b300-1028a1/","markdown":"https://newruntime.com/posts/hyperstition-open-sources-hyperkernel-megakernels-for-nvidia-b300-1028a1.md","json":"https://newruntime.com/posts/hyperstition-open-sources-hyperkernel-megakernels-for-nvidia-b300-1028a1.json"}
markdown_url: "https://newruntime.com/posts/hyperstition-open-sources-hyperkernel-megakernels-for-nvidia-b300-1028a1.md"
---

# Hyperstition open-sources Hyperkernel megakernels for NVIDIA B300

## Retrieval answer

Hyperstition has published an MIT-licensed alpha of Hyperkernel, a set of megakernels and command-line serving paths for NVIDIA B300 GPUs. It spans LLM inference, driving, DNA generation, video world models, robotics, and protein folding, giving equipped teams public code to inspect and test rather than a turnkey service or demonstrated performance advantage.

## Summary
Hyperstition has published an MIT-licensed alpha of Hyperkernel, a set of megakernels and command-line serving paths for NVIDIA B300 GPUs. It spans LLM inference, driving, DNA generation, video world models, robotics, and protein folding, giving equipped teams public code to inspect and test rather than a turnkey service or demonstrated performance advantage.

## What happened
Hyperstition’s public GitHub repository describes Hyperkernel as the public implementation of megakernels generated by the project. It includes Hyperserve, a command-line interface that selects a workload-specific serving or execution path. The repository is MIT-licensed and explicitly labelled alpha. Its README says it will be updated over the next few weeks with more models, API simplifications, and performance improvements.

The current kernels are compiled for NVIDIA B300 GPUs. The repository lists four kernels for LLMs, one each for driving, DNA, robotics, and protein folding, plus two for world models. It does not bundle any model weights. Instead, an operator supplies checkpoint directories containing safetensors files, and some workloads require conversion scripts before those weights can be used. Building the project requires Rust, according to the repository’s toolchain configuration, and an NVIDIA R580 driver or newer.

For language-model serving, Hyperserve runs DeepSeek-V4.1-Flash as an OpenAI-compatible server for chat or completion JSON requests. This path uses four GPUs. The command accepts a checkpoint directory, host and port options, and an optional API key; when configured, requests must present a bearer token. The README lists presets from c1 through c8 for maximum concurrency, while noting that c6 and c7 use the c5 kernels. All listed presets have a 1 million-token context. The LLM configuration also requires about 203 GB of host RAM for Engram tables, which are mapped once across the four GPU processes.

The non-LLM commands are more narrowly shaped around supplied model inputs. The driving command runs the openpilot v0.11.2 driving model over camera frames on one GPU and writes outputs per frame. The DNA path uses Evo 2 7B on one GPU, accepts a FASTA or plain-text prompt sequence, and permits prompt plus generation steps within 1,048,576 positions. The world-model command runs Matrix-Game 3.0 on two GPUs and generates an initial 57-frame clip followed by optional 40-frame chunks. Its action input is a text file with one control line per frame.

Robotics support includes ACT and pi0-FAST paths, each on one GPU. ACT takes camera images and optional joint states, returning chunks of 14-joint actions; the README says its ResNet-18 backbone, transformer encoder and decoder, action head, and normalization are executed in one kernel launch. The pi0-FAST mode processes three camera inputs, token IDs, and prefix lengths to generate action chunks. A fold command is listed for ESMFold. Across these modes, the README specifies data formats, conversion steps, and command options, but it does not include a hosted endpoint, packaged checkpoints, or a managed deployment path.

The release is consequently a public implementation and inspection point for selected B300 execution paths. The supplied material does not provide benchmark tables, latency, throughput, utilization, quality, or cost comparisons. It also does not claim production readiness: Hyperkernel is described as alpha and its maintainers explicitly place further model support, API work, and performance changes in future updates.

## What the archive adds
This is a new fact on a related infrastructure thread. [Nvidia's GPU-sharing software is now free and open source](https://newruntime.com/posts/nvidia-s-gpu-sharing-software-is-now-free-8c9a45/) concerned orchestrating GPU workloads, whereas Hyperkernel exposes workload-specific kernels and execution commands. [NVIDIA Nemotron 3.5 Bolt E4B small open model becomes available on Ollama](https://newruntime.com/posts/nvidia-nemotron-3-5-bolt-e4b-small-open-c12f6c/) was a model-distribution event, not prior coverage of Hyperkernel.

## Trend signal
This weakly supports [harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/). Hyperkernel puts multiple model and task families behind one project with distinct execution paths, implying that packaging, input handling, kernel selection, and service interfaces can be considered separately from a single model identity. One alpha repository is not evidence that this boundary is durable, widely adopted, or portable beyond B300 hardware.

## Why it matters
For teams with B300 access, Hyperkernel provides a concrete artifact to examine rather than a high-level claim about specialized inference. Engineers can inspect how a project separates model checkpoints from kernels, what host-memory and GPU assumptions it makes, and how a single command surface is adapted to radically different inputs: chat requests, camera frames, DNA sequences, video actions, and robot observations.

The breadth of the listed workloads is notable. Most serving discussions center on text models, while this repository also names driving, robotics, world models, and biological workloads. That does not prove common infrastructure across those domains is operationally mature. It does give systems teams an example of a project attempting to package those execution modes together. The LLM server’s OpenAI-compatible request format may also make it easier to test that particular path against clients that already speak that interface.

Its practical applicability is sharply bounded. A team needs B300 GPUs, the appropriate driver, locally obtained and sometimes converted checkpoints, and tolerance for alpha software. Four GPUs and approximately 203 GB of host RAM for the LLM mode make it unsuitable as a general local-serving recipe. Without published measurements, no builder should infer a speed, cost, quality, or operational advantage over another serving stack.

## Limits and uncertainty
The repository establishes the existence of code, named workloads, hardware requirements, supported checkpoint references, and invocation details. It does not establish measured latency, throughput, memory efficiency beyond the stated LLM host-memory requirement, output quality, or cost per completed task. There are no supplied comparative results against vendor software or other open-source serving systems.

The material also does not show independent validation, production load behavior, failure handling, observability, security review, or maintenance commitments. Different commands may be at different levels of completeness; the workload list alone is not evidence of equivalent support. Since weights are excluded and some paths need conversion tooling, reproducibility depends on external checkpoints and preparation steps not delivered by this repository.

## Reader takeaway
Hyperkernel is an MIT-licensed, alpha-stage B300 project that publishes specialized kernels and Hyperserve command paths for several AI workloads. It is worth examining if your team already operates compatible hardware and can prepare the required checkpoints. Treat it as an implementation to test and learn from, not as evidence of a proven production-serving or cost advantage.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Inference](https://newruntime.com/topics/inference.md): Explore the inference topic hub.
- topic hub: [Compute](https://newruntime.com/topics/compute.md): Explore the compute topic hub.
- topic hub: [Open Source](https://newruntime.com/topics/open-source.md): Explore the open-source topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)