HyperNova 60B: Multiverse's Compressed gpt-oss-120b Model
HyperNova 60B is an open-weight language model from Multiverse Computing, a Spanish company that compresses other labs' models. It takes OpenAI's gpt-oss-120b and shrinks its weights from 65 GB to 32 GB with a method the company calls quantum-inspired. TechCrunch reported on February 24, 2026 that the updated 2602 version was free to download on Hugging Face, and a 2605 version followed. Below you will find how the compression works and what it costs in benchmark points. The last sections cover when that cost is worth paying.
What is HyperNova 60B, in one sentence?
HyperNova 60B is a mixture-of-experts model with about 60 billion total and 4.8 billion active parameters, made by compressing gpt-oss-120b with Multiverse's CompactifAI method and released under Apache 2.0.
The model card lists it as a decoder-only Transformer with native tool calling in OpenAI's function-calling format. Reasoning effort is configurable at low, medium or high in the system prompt. "Quantum-inspired" refers to the math behind the compression. The model runs on ordinary GPUs and needs no quantum computer.
How does HyperNova 60B compare with gpt-oss-120b?
Multiverse publishes its own comparison on the model card. Every number below is the vendor's, run on one NVIDIA H200 at high reasoning effort.
| gpt-oss-120b | HyperNova 60B 2602 | HyperNova 60B 2605 | |
|---|---|---|---|
| Weights | 65 GB | 32 GB | 32 GB |
| Total parameters | 117B | 59B (4.8B active) | 59B (4.8B active) |
| MMLU-Pro | 79.6 | 74.3 | 76.8 |
| GPQA Diamond | 74.6 | 65.6 | 71.9 |
| LiveCodeBench | 62.8 | 51.5 | 68.7 |
| Tau2-bench Telecom | 63.7 | 60.5 | 61.7 |
| Terminal Bench | 24.2 | 12.1 | 15.9 |
| Humanity's Last Exam | 18.5 | 7.3 | 15.0 |
Sources: the HyperNova 60B 2605 model card and Multiverse's 2602 technical article. The gpt-oss-120b parameter count comes from the same article.
The 2605 version stays within a few points of the original on knowledge and tool-use tests and beats it on LiveCodeBench in Multiverse's run. It still loses clearly on Terminal Bench and Humanity's Last Exam. Those two are the tests closest to long agentic work and hard reasoning.
How does CompactifAI compress a model?
Most compression removes parts of the network (pruning, distillation) or lowers each weight's precision (quantization). The CompactifAI paper, first posted in January 2024, takes another route. It rewrites weight matrices as tensor networks, a factorization borrowed from quantum physics, to compress what the authors call the model's correlation space. The method can be stacked with quantization.
For HyperNova, Multiverse's technical article says the decomposition targeted redundancy in gpt-oss-120b's expert routing structure. A knowledge-distillation phase followed. A larger teacher model generated thousands of synthetic examples to retrain tool calling and multi-turn handling. Per the same article, that step lifted the 2602 version's BFCL v4 tool-calling score from 25 to 62. Multiverse says the result recovers 97% of the base model's tool-calling capability.
What hardware does HyperNova 60B need?
Multiverse's article says the model "fits on a single 40 GB GPU," and that an 80 GB card can hold two instances. For serving, the card shows vLLM with an OpenAI-compatible API, plus SGLang and Docker Model Runner. Hugging Face Transformers also works.
On speed, Multiverse reports the 2605 version at 5,210 tokens per second against 3,821 for gpt-oss-120b. Time to first token was 4.85 seconds against 7.04. That run used an H200 at a concurrency of 128 with 1,000 input and 1,000 output tokens per request. Those figures come from a datacenter GPU. A 32 GB model is also within reach of a well-equipped workstation, but Multiverse publishes no laptop numbers.
When should you use a compressed model?
A compressed model makes sense when GPU memory is the constraint. HyperNova fits on a 40 GB card, per Multiverse, which can cut serving cost. Like any open-weight model, it also lets you keep data on hardware you control.
It is a worse fit for agent loops that run many terminal steps. HyperNova 2605 scores 15.9 on Terminal Bench against 24.2 for the original. On a long task, a lower success rate per step compounds with every step.
Read the fine print
Every benchmark here is from Multiverse. We found no independent evaluation of the 2605 version. Multiverse's claim to TechCrunch that the model beats Mistral Large 3 is also unverified in that article. The card's headline claims favor the 2605 version's strengths, but its weakest results (Terminal Bench and Humanity's Last Exam) are in the same table.
HyperNova keeps gpt-oss-120b's architecture. The card gives no separate context length for HyperNova. Its Terminal-Bench setup used a maximum model length of 131,072 tokens.
The bottom line
HyperNova 60B is worth a test if you serve gpt-oss-120b today and GPU memory is the bill you want to cut. Keep the full model for long agentic coding until the Terminal Bench gap closes. Treat every number on the card as a starting hypothesis for your own evaluation.
Does agent memory survive a switch to a smaller model?
You might move a workload from gpt-oss-120b to HyperNova to save memory, then back when a task needs more reasoning. What your agent learned about your project last week lives in session history and instruction files. A new model starts without it unless it is stored outside the model.
memU stores that knowledge as readable Markdown, outside any model. A scheduled background task slices new session logs into jobs. The agent distills each job into memory and skill files, which memU commits to one backend shared across hosts. memU's core makes no LLM calls. It indexes the files the agent wrote. A standing instruction in CLAUDE.md or AGENTS.md tells the agent to retrieve them before answering, whichever model the host calls next. The self-hosted option keeps memory on one private device and needs an embedding key.
To try the hosted version with a key from memu.so, send your agent one message:
Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU
The self-hosted install and supported hosts are in the memU GitHub repository. The reasoning behind the design is in why memU gives memory back to the agent.
FAQ
Is HyperNova 60B free?
Yes. The weights are on Hugging Face under the Apache 2.0 license, which allows commercial use.
Is HyperNova 60B a quantum model?
No. "Quantum-inspired" describes the tensor-network math used to compress it. The model runs on standard GPUs.
How big is HyperNova 60B?
The weights are about 32 GB, half the 65 GB of gpt-oss-120b. Multiverse says it fits on a single 40 GB GPU.
What is HyperNova 60B based on?
It is a compressed version of OpenAI's open-weight gpt-oss-120b. Multiverse applied its CompactifAI tensor-network compression and then retrained tool calling with knowledge distillation.
Can HyperNova 60B call tools?
Yes. The model card lists native tool calling in OpenAI's function-calling format, plus structured outputs for agent workflows.