OrcaRouter releases OrcaCyber Nano 1.0, a 27B local security model with 262K context and GGUF builds
OrcaRouter has published OrcaCyber Nano 1.0, a 27B vision-language reasoning model for defensive security, as GGUF files for local inference with llama.cpp. It supports a 262,144-token context window and text, image and video input. Benchmark scores are self-reported, and the harness details are not yet published.
OrcaCyber Nano 1.0 — Quick Specs
OrcaRouter has released OrcaCyber Nano 1.0, a 27B-parameter vision-language reasoning model for defensive security work, as GGUF files that run locally through llama.cpp. The model has a 262,144-token context window and takes text, image and video input. Output is text only.
Access is gated on Hugging Face. Users must accept conditions and share their Hugging Face information to download the files. The repository reports 19,697 downloads in the last month.
Model details
- Parameters: 27B
- Architecture:
Qwen3_5ForConditionalGeneration, a hybrid vision-language design - Layers / hidden size: 64 layers, 5120 hidden; 48 linear-attention and 16 full-attention layers
- Context window: 262,144 tokens
- Modalities: text, image and video in; text out
- Base model:
orcarouter/OrcaCyber-Nano-1.0-AI-Defense; this repo is the quantized release - Training cutoff: not disclosed
- Pricing: no API pricing disclosed; the weights are downloadable under gated access
Orca says the model targets reverse engineering and decompilation, vulnerability analysis, secure code review, triage of logs, alerts and packet captures, and incident response. The vision projector is meant for inputs such as Ghidra or IDA screenshots and network diagrams. Orca says the model keeps general-purpose reasoning and can also serve as a local coding assistant.
Files and serving
The repo ships four files, and Orca says all load in stock llama.cpp with no custom kernels:
| File | Size | Role |
|---|---|---|
| AI-Defense-Mix-15.7G.gguf | 15.70 GB | Best-quality build, built-in MTP head |
| AI-Defense-Mix-13G.gguf | 12.99 GB | Fits a 16 GB card with the MTP drafter |
| AI-Defense-DFlash2-Q8_0.gguf | 2.06 GB | DFlash2 speculative drafter |
| mmproj-...-F16.gguf | 0.93 GB | Vision projector for image input |
Speculative decoding comes in two forms: a multi-token-prediction (MTP) head built into both main files, enabled with --spec-type draft-mtp, and a separate DFlash2 drafter. The 2B parameter figure on the page's metadata panel refers to the drafter file, not the main model. Recommended sampling is temperature 1.0, top_p 0.95 and top_k 20.
Benchmarks (self-reported)
Orca lists these scores. It says the harness, settings and full result tables are "being finalized," so none are independently verified.
- MATH-500: 98.60%
- GSM8K: 97.50%
- AIME 2025: 96.23%
- HumanEval+: 93.29%
- LiveCodeBench: 92.23%
- GPQA Diamond: 88.28%
- MBPP+: 84.66%
- IFEval: 82.81%
- SWE-bench: 78.4%
- Terminal-Bench 2.1: 72.5%
- IFBench: 56.33%
- BigCodeBench: 49.82%
The page does not say which SWE-bench variant was used. It lists no cybersecurity-specific benchmark for Nano 1.0. The 98.01% pass@1 on CyberGym Level 1 on the same page belongs to a different model, OrcaCyber Zero 1.5, which is also an Orca claim.
What this means
The practical appeal is deployment. A 27B model that fits a 16 GB card at 12.99 GB, with speculative decoding built in, lets security teams analyze proprietary binaries and unreleased vulnerabilities on air-gapped machines. That addresses a real constraint for the target users.
The benchmark table needs caution. Scores such as 96.23% on AIME 2025 and 98.60% on MATH-500 at this size are unusually high and come with no published harness. The model's general reasoning scores also say little about security tasks, and no security-specific results are given for Nano 1.0. Until Orca publishes evaluation settings or third parties reproduce the results, treat the numbers as marketing claims. The licensing terms behind the gate also matter for commercial use, and they are not visible without accepting the conditions.
Related Articles
NVIDIA releases Nemotron 3.5 Super VL NVFP4: 120B-parameter, 12B-active vision-language model with 1M context
NVIDIA has published Nemotron 3.5 Super VL in NVFP4 quantization on Hugging Face: a 120B-parameter, 12B-active hybrid Mamba-2/Transformer MoE with a 1M-token context window and text, image and video input. NVIDIA reports the quantized checkpoint stays close to BF16 on most benchmarks. It fits on one DGX Spark, one H200, two H100s or one GB200.
IBM Granite Speech 5.0 2B ranks #2 among open ASR models at 5.06% WER under Apache 2.0
IBM released granite-speech-5.0-2b, a 2B-parameter speech-language model trained on 390,000 hours of public audio. IBM claims it ranks #2 among open-weight ASR models with a 5.06% average English WER. IBM also updated its 470M Turbo CTC model to 19,870 RTFx and refreshed the weights of granite-speech-4.1-2b-plus.
NVIDIA releases Kumo Forecast, an Apache 2.0 time-series model built on a ~340M-parameter MOMENT-1-large encoder
NVIDIA has released Kumo Forecast (formerly NV-Tesseract Forecasting), a multivariate time-series forecasting model under Apache 2.0. It pairs a frozen MOMENT-1-large encoder of roughly 340 million parameters with a trainable forecasting head, optional nearest-neighbor retrieval over historical data (DARR), and a built-in interpretability framework.
NVIDIA releases Nemotron 3.5 Embed 8B, scoring 61.17 nDCG@10 on ViDoRe v3.1 among dense retrievers
NVIDIA has released Nemotron 3.5 Embed (8B BF16, 1B BF16, 1B NVFP4), Nemotron 3.5 Rerank 8B BF16 and Nemotron 3.5 Parse (631M parameters) as open, commercially available models. NVIDIA reports that the 8B embedder ranks first among dense retrievers on ViDoRe v3.1 with 61.17 nDCG@10, and that adding the reranker cuts agent token cost per query by roughly 13-14%.
Comments
Loading...