luxopes/Cognix-1-Flash
Cognix-1-Flash is a 1.2 billion parameter, two-stage merged BF16 model developed by LuxAI, based on the LFM2.5-1.2B architecture. It was initially trained with a code-first SFT epoch and further refined with specific identity and refusal examples. This model is experimental and not intended as a reliable reasoning assistant, with measured low accuracy across math, code, and reasoning tasks, and a known identity issue.
Loading preview...
Model Overview
Cognix-1-Flash is a 1.2 billion parameter, two-stage merged BF16 model developed by LuxAI. It is built upon the LFM2.5-1.2B base architecture and features a 32768 token context cap. The model underwent an initial code-first supervised fine-tuning (SFT) stage, followed by a second stage focusing on identity, refusal, and benign-adjacent examples, including deterministic original code replay.
Key Characteristics
- Architecture: Two-stage merged BF16 model based on LFM2.5-1.2B.
- Parameter Count: 1.2 billion parameters.
- Context Length: Supports a 32768 token context cap, though long-context performance is not validated.
- Training Focus: Initial code-first SFT, followed by refinement with identity, refusal, and helpful-cyber examples.
- Tokenizer: Uses the unchanged native tokenizer and special tokens.
- Availability: Provided as a full BF16 model, with GGUF downloads available in F16 and Q8_0 formats.
Measured Limitations
This model is explicitly described as experimental and not a reliable reasoning assistant. Benchmarks on 240 Czech Lux Core 1 tasks (Math, Code, Knowledge, Reasoning, Reading, Instructions) showed an overall accuracy of 21.67%. Specific limitations include frequent arithmetic and logic mistakes, and a known identity issue where the model falsely claimed OpenAI authorship without an identity-setting system prompt. No quality or safety guarantees are claimed.