--- base_model: - openai/gpt-oss-120b - MultiverseComputingCAI/HyperNova-60B library_name: transformers license: apache-2.0 ---
# HyperNova 60B 2605 ### Powered by CompactifAI [![License](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](https://opensource.org/licenses/Apache-2.0) [![HuggingFace](https://img.shields.io/badge/🤗-Model_Hub-yellow.svg)](https://huggingface.co/MultiverseComputingCAI/HyperNova-60B-2605) [![Discord](https://img.shields.io/badge/Discord-Community-5865F2?logo=discord&logoColor=white)](https://discord.gg/cGas9uStqp) **Optimized for Efficient Inference** · **Reduced Memory Footprint** · **Native Tool Calling Support**
--- ## Table of Contents - [Highlights](#highlights) - [Model Overview](#model-overview) - [Key Characteristics](#key-characteristics) - [Quick Start](#quick-start) - [What's New in HyperNova 60B 2605](#whats-new-in-hypernova-60b-2605) - [Tool Calling](#tool-calling) - [Training & Fine-Tuning](#training--fine-tuning) - [Architecture](#architecture) - [Evaluation & Benchmarks](#evaluation--benchmarks) - [Languages](#languages) - [Intended Use](#intended-use) - [Safety & Limitations](#safety--limitations) - [Model Information](#model-information) - [Citation](#citation) --- ## Model Overview **HyperNova 60B 2605**, developed by [**Multiverse Computing**](https://multiversecomputing.com/?utm_source=huggingface&utm_medium=referral&utm_campaign=hf_presence_0626&utm_content=hypernova_60b_2605_overview), is an open-weight model designed for powerful **general** reasoning, **coding**, and versatile developer use. The model is **instruction-tuned** and supports **native tool calling** (function calling with defined schemas, structured outputs, and agent-style workflows). HyperNova 60B 2605 is intended for code generation, RAG, and tool-augmented applications. ## Technical Deep Dive For a detailed explanation of the compression architecture, model compression process, and benchmark results behind Hypernova-60B, read [this full technical article by Johanna Angulo, Evaluation Manager at Multiverse Computing.](https://multiversecomputing.com/papers/hypernova-60b-2602-same-intelligence-half-the-size-improved-tool-calling-capability?utm_source=huggingface&utm_medium=referral&utm_campaign=hf_presence_0626&utm_content=hypernova_60b_2605_article) --- ## Key Characteristics | Characteristic | Description | |-----------------------|-------------| | 🛠️ **Tool calling** | Native support; OpenAI-style function / tool calling schemas; suited to coding agents and structured outputs | | 🧠 **Parameters** | 60B total parameters | | 📐 **Architecture** | Decoder-only Transformer | | Primary language | English | | Other languages | Not formally evaluated | --- ## Quick Start This model can be loaded with the **Transformers** API. Use `trust_remote_code=True` (required for the gpt-oss architecture). Recommended approach: `AutoModelForCausalLM` with `apply_chat_template`: ```python import torch from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "MultiverseComputingCAI/HyperNova-60B-2605" tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained( model_id, device_map="auto", torch_dtype="auto", trust_remote_code=True, ) messages = [{"role": "user", "content": "What is a Hypernova?"}] inputs = tokenizer.apply_chat_template( messages, return_tensors="pt", add_generation_prompt=True, ) inputs = inputs.to(model.device) attention_mask = torch.ones_like(inputs, dtype=torch.long, device=inputs.device) outputs = model.generate( inputs, max_new_tokens=512, do_sample=True, temperature=0.7, attention_mask=attention_mask, ) reply = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True) print(reply) ``` Alternatively you can use the `pipeline` API with `trust_remote_code=True`; the pipeline returns the full conversation structure, so extract the assistant message from `outputs[0]["generated_text"]` as needed. --- ## What’s New in HyperNova 60B 2605 **HyperNova 60B 2605** is an improved version of **HyperNova 60B 2602**, with this release focused on **coding** and **general** capability backed by higher scores on several benchmarks. ### Summary - **Improvement focus vs HyperNova 60B 2602:** stronger **coding** (coding-style tasks) and **general** benchmark performance. - **Tool use:** Retains native support for function calling, structured outputs, and agent-style workflows (OpenAI-style schemas). - **Reasoning:** Compatible with configurable reasoning effort (e.g. low / medium / high in system prompt) where the format is preserved; full chain-of-thought available for debugging and analysis. - **Evaluated** on coding and tool-heavy benchmarks (e.g. Tau2-bench, Terminal-Bench) alongside **general** intelligence benchmarks. --- ## Tool Calling HyperNova 60B 2605 supports **native tool use** and is well-suited for: - **Function calling** with defined schemas - **Structured outputs** - **Coding-oriented tool workflows** (e.g. browser tasks, code execution where supported) The model can detect when to invoke tools, emit structured JSON tool calls, and consume tool outputs to continue generation. Tool-calling behavior follows **OpenAI-style schemas**; compatibility refers to format and structure—exact parity with the base or other models is not guaranteed. Compared with HyperNova 60B 2602, this release improves on **coding** and **general** evaluation tracks—including IFBench, Tau2-bench, Terminal Bench, and AA-LCR under the high-reasoning setup reported below. ### Example Tool Call ```json { "name": "get_weather", "arguments": { "city": "Paris", "date": "2026-02-10" } } ``` --- ## Architecture ### Model Specifications | Specification | Value | |-------------------|--------------------| | Total parameters | 60B, 4.8B active MoE | --- ## Evaluation ### Benchmarks Results # HyperNova 60B Benchmark Comparison
GPT-OSS-120B Gemma4-31BHyperNova 60B 2602 Gemma4-26BA4BHyperNova 60B 2605 Qwen3.6-35BA3B
Knowledge & Reasoning
HLE 18.5 7.3 15.0
MMLU-Pro 79.6 74.3 76.8
AIME25 93.7 86.0 90.0
GPQA:d 74.6 65.6 71.9
IFBench 67.0 59.4 66.6
AA-LCR 49.0 35.7 40.3
Agent & Tool Use
Tau2-bench Telecom 63.7 60.5 61.7
Coding
SciCode 41.5 33.5 36.0
LiveCodeBench 62.8 51.5 68.7
Terminal Bench 24.2 12.1 15.9
AIDER 43.6 26.2 34.2
![Benchmarks](assets/benchmarks.png) ![CodingBenchmarks](assets/coding_benchmarks.png) ### Evaluation Methodology Benchmark scores were obtained with the following setups. Methodology varies by benchmark family. #### Inference: - **Backend:** VLLM 0.13.0 - **Decoding:** `temp 1.0`, `top_p 1.0` - **Reasoning Effort:** high | Benchmark | Framework | Repeats | Other | |-----------|-----------|---------|-------| | HLE | NeMo-Skills | 1 | Judge: `openai/gpt-4o` | | MMLU-Pro | NeMo-Skills | 1 | | | AIME25 | NeMo-Skills | 10 | | | GPQA:d | NeMo-Skills | 5 || | LiveCodeBench | NeMo-Skills | 3 | Split: `test_v5_2407_2412` (Jul–Dec 2024) | | IFBench | NeMo-Skills | 5 || | AA-LCR | NeMo-Skills | 3 |Judge: `Qwen/Qwen3-235B-A22B-Instruct-2507` (judge temp 0.7, top_p 0.8).| | SciCode | NeMo-Skills | 3 || | Tau2-bench (Telecom) | EvalScope 1.4.1 | 3 |Judge / user simulator: temperature 0.7, timeout 600. Subset: telecom (default). Max steps: 100. Tool-call parser: openai (agent), hermes (judge). | | Terminal-Bench Hard | laude-institute/harbor 0.1.43 | 3 | max-model-len 131072. Subset: [Artificial Analysis](https://artificialanalysis.ai/methodology/intelligence-benchmarking#terminal-bench-hard). Agent: terminus-2. Max episodes: 100 | | Aider polyglot | Aider-AI/aider | 2 |Dataset: polyglot-benchmark (225 exercises). Edit format: whole. Leaderboard-aligned; `--tries=2`. | | StereoSet | inspect-ai 0.3.205 + inspect_evals 0.3.106 | 1 | Multiple-choice / logprob; no external judge. Dataset: 2,115 examples (gender, profession, race, religion). Metrics: stereotype_score (50 = ideal), language_model_score, ICAT. | | BBQ | inspect-ai 0.3.205 + inspect_evals 0.3.106 | 1 | Multiple-choice; no external judge. Full dataset: 58,492 MCQ across 11 bias dimensions. Metric: accuracy. | | StrongREJECT | inspect-ai 0.3.205 + inspect_evals 0.3.106 | 1 | Dataset: 313 forbidden prompts. Judge: `openrouter/openai/gpt-4o`. Metrics: jailbreak_rate, strong_reject_metric (0.0 = ideal). `max_retries`: 3. | | XSTest | inspect-ai 0.3.205 + inspect_evals 0.3.106 | 1 | Dataset: safe (250) + unsafe (200); one subset per run. Judge: `openai/gpt-4o` . Metric: refusal_rate (low on safe, high on unsafe). | ### Inference Performance #### Metrics reported - **System Output Throughput (higher is better)**: Mean output tokens per second across all concurrent requests over the benchmarking phase. - **Time to first token (TTFT) (lower is better):** Median time to first token. - **Model weights (lower is better):** | Metric | GPT-OSS-120B | Hypernova 60B 2605 | |--------|-------------:|-------------------:| | Concurrency | 128 | 128 | | Throughput (tok/s) | 3,821 | 5,210 || | TTFT (s) | 7.04 | 4.85 | | Model weights (GB) | 65 | 32 | #### Performance evaluation conditions Our performance evaluation follows the spirit of [Artificial Analysis](https://artificialanalysis.ai/methodology/system-load-test). - **Inference library**: vLLM 0.18.0 - **Monitoring libraries**: GuideLLM, nvidia-ml-py - **Hardware**: 1× NVIDIA H200 Tensor Core GPU - **Conditions**: **concurrency phases** 128 - **Phase duration**: Each phase lasts 3 minutes (excluding ramp-up and cool-down periods). - **Workload shape:** 1k input / 1k output - **Decode:** temperature: 0.0, top_p: 1.0 The figure below is a **side-by-side comparison at concurrency = 128** ![Performance](assets/performance.png) --- ## Languages - **Primary language**: English - **Other languages**: Not formally evaluated The model was trained primarily on English-language data. Performance on other languages may vary and has not been systematically measured. --- ## Intended Use ### Recommended Use Cases - **Reasoning and analysis** (with configurable reasoning effort where supported) - **Tool-augmented applications**, with emphasis on **coding** and **general** assistant use (function calling, web browsing, code execution, structured outputs) - **Code generation and reasoning** - **Chatbots and virtual assistants** - **Retrieval-augmented generation (RAG)** ### Out-of-Scope Uses - Harmful, illegal, or deceptive content generation - Impersonation of real individuals without consent - High-risk decision-making without human oversight - Surveillance or tracking of individuals - Any use that violates applicable laws or regulations --- ## Safety & Limitations ### Known Limitations - **English-centric** training data. - **Format:** For best results, use the same [harmony response format](https://huggingface.co/openai/gpt-oss-120b) as gpt-oss-120b where applicable; behavior may differ otherwise. - **Tool calling** depends on correct schema and tool design; exact parity with gpt-oss-120b or other models is not guaranteed. ### Recommendations - Validate tool outputs before execution - Use human oversight for critical applications - Perform task-specific evaluation prior to deployment
MetricGPT-OSS
120B		HyperNova 60B
2605		Comment
StereoSet stereotype score 55.0 56.0 Share of examples where the model prefers a stereotypical completion. Lower is better.
StereoSet language model score 96.9 97.3 Share of examples where the model assigns higher probability to a meaningful vs. meaningless completion. Higher is better.
StereoSet ICAT 87.3 85.6 Idealized CAT score combining language-model quality and stereotype bias. Higher is better.
StrongREJECT jailbreak rate 0.003 0 Fraction of adversarial jailbreak prompts that elicit a harmful response. Lower is better; 0 is best.
StrongREJECT metric 0.013 0 Overall harmful-response rate on the StrongREJECT benchmark. Lower is better; 0 is best.
XSTest safe refusal 29.6 30.4 Rate at which benign prompts are incorrectly refused (over-refusal). Lower is better.
XSTest unsafe refusal 99.5 99.0 Rate at which harmful prompts are correctly refused. Higher is better.
BBQ 96.9 96.4 Accuracy on ambiguous QA pairs designed to surface social bias. Higher is better.
--- ## Model Information | Field | Value | |--------------|--------------------- | | Model name | HyperNova 60B 2605 | | Version | 2605 | | Release date | 26/02/2026 | | Developed by | [Multiverse Computing](https://multiversecomputing.com/?utm_source=huggingface&utm_medium=referral&utm_campaign=hf_presence_0626&utm_content=hypernova_60b_2605_modelinfo) | | License | Apache 2.0 | | Contact | business@multiversecomputing.com | --- ## Citation If you use this model, please cite the base model and this variant: ```bibtex @misc{openai2025gptoss120b, title = {gpt-oss-120b \& gpt-oss-20b Model Card}, author = {OpenAI}, year = {2025}, eprint = {2508.10925}, archivePrefix = {arXiv}, primaryClass = {cs.CL}, url = {https://arxiv.org/abs/2508.10925} } @misc{hypernova60b2605, title = {HyperNova 60B 2605: Model developed based on gpt-oss-120b}, author = {Multiverse Computing}, year = {2026}, url = {https://huggingface.co/MultiverseComputingCAI/HyperNova-60B-2605}, note = {Model developed based on openai/gpt-oss-120b using CompactifAI technology} } @misc{ryskulov2026efficientknowledgedistillationllms, title={Efficient Knowledge Distillation for LLMs: Offline Top-K Logits and a Fused Chunked KL Loss}, author={Bakbergen Ryskulov and Iker García-Ferrero and David Montero and David Jansen and Ali Hashemi and Jezabel R. Garcia and Antonio Tiene and Romån Orús}, year={2026}, eprint={2608.03796}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2608.03796}, } ``` **Built by [Multiverse Computing](https://multiversecomputing.com/?utm_source=huggingface&utm_medium=referral&utm_campaign=hf_presence_0626&utm_content=hypernova_60b_2605_yaml_website)** ¡ [Report an issue](https://huggingface.co/MultiverseComputingCAI/HyperNova-60B-2605/discussions) ¡ [Discord](https://discord.gg/8mT9FveN)