Strengths
- Permissive Apache 2.0 license, with weights in three formats (BF16, FP8 and NVFP4) on Hugging Face.
- Reflection AI reports 81.6% on the SWE-bench Verified coding test and 80.1% on Terminal Bench 2.1, higher than NVIDIA reports for Nemotron 3.5 Super VL.
- A 1 million token context window, so it can read large code projects or long documents in one go.
- Five reasoning levels from low to max, so you can trade speed for thoroughness.
- Wide availability: AWS, Google Cloud, Microsoft Foundry, Oracle, Nebius, Together, Cloudflare, OpenRouter and on-premises servers through Dell and HPE.
- Cheap API from providers: from $0.25 per million input tokens and $1 per million output tokens.
Weaknesses
- It understands text only. It cannot handle images, scans or screenshots.
- Reflection AI built it mainly for English and lists no other languages on the model page. Test it before using it for other languages.
- Running it yourself takes a lot of hardware: even the smallest NVFP4 version is about 280 GB, and the recommended setup is four B200 GPUs.
- All test results come from the vendor. It is the first open model from a young company, so there is little independent experience with it so far.
- Reasoning cannot be switched off, only lowered to the low level. Answers therefore take longer and use more tokens, and OpenRouter defaults to the max level.
Best for
- AI coding agents
- long multi-step agent tasks
- working with large codebases or long documents
- your own model in a company data center
Not ideal for
- images and scanned documents
- small servers and ordinary PCs
- texts in Slovak or Czech without testing first
How to use it in Europe
The weights are free to download from Hugging Face anywhere in Europe, and the Apache 2.0 license has no geographic restrictions. On your own servers your data stays in house, but you need powerful GPUs such as B200 or H200. Through an API, providers such as Nebius, Together and CoreWeave offer it via OpenRouter, as do big clouds like AWS and Google Cloud. The US company Reflection AI runs its own API as a free beta, and its docs do not yet describe data retention. For sensitive data, check where your provider processes it.
Price compared with similar models
Benchmark results
| Benchmark | Score | Source |
|---|---|---|
| SWE-bench Verified | 81.6% | huggingface.co ↗ (vendor-reported), October 2026 |
| Terminal Bench v2.1 | 80.1% | huggingface.co ↗ (vendor-reported), October 2026 |
| GPQA Diamond | 90.6% | huggingface.co ↗ (vendor-reported), October 2026 |
| HLE with tools | 51.1% | huggingface.co ↗ (vendor-reported), October 2026 |
| BrowseComp with context management | 76.8% | huggingface.co ↗ (vendor-reported), October 2026 |
Benchmarks cover only a narrow slice of ability and vendors often report them at their best settings. Treat them as a rough guide.
Similar models from other vendors
- Claude Opus 5.5 (Anthropic): Anthropic's main model for demanding work, coding and long documents, and the one Anthropic recommends starting with.
- DeepSeek V4 Pro (DeepSeek): DeepSeek's largest open-weight model for demanding text tasks at a low price. It does not understand images.
- Gemini 4 Carbon (Google): Google's most capable model, announced on 28 October 2026. So far only selected cybersecurity experts have access.
- Gemini 4 Argon (Google): Google's most capable publicly available model, out in October 2026. Built for demanding analytical work where accuracy matters and made-up answers are costly.
- Muse Wonder (Meta): Meta's strongest model for long-running agents and coding, still released as a preview.
- Mistral Large 4 (Mistral AI): Mistral's largest model, still in preview. Built for demanding tasks, long documents and security analysis with data kept in the EU.