| { |
| "model": "Qwen/Qwen3-Coder-Next", |
| "arch": "Qwen3-Coder-Next (80B MoE, 3B active, 48L, 512 experts)", |
| "max_position_embeddings": 262144, |
| "tfa_layers": [ |
| 3, |
| 11, |
| 23, |
| 35, |
| 47 |
| ], |
| "tfa_config": { |
| "heads": 8, |
| "head_dim": 128, |
| "hidden": 2048 |
| }, |
| "kv_cache_benchmark": [ |
| { |
| "seq_len": 128, |
| "tfa_state_MiB": 2.5, |
| "status": "proved_prior_run" |
| }, |
| { |
| "seq_len": 512, |
| "tfa_state_MiB": 2.5, |
| "status": "proved_prior_run" |
| }, |
| { |
| "seq_len": 1024, |
| "tfa_state_MiB": 2.5, |
| "status": "proved_prior_run" |
| }, |
| { |
| "seq_len": 4096, |
| "tfa_state_MiB": 2.5, |
| "status": "proved_prior_run" |
| }, |
| { |
| "seq_len": 16384, |
| "tfa_state_MiB": 2.5, |
| "status": "proved_prior_run" |
| }, |
| { |
| "seq_len": 65536, |
| "tfa_state_MiB": 2.5, |
| "status": "proved_prior_run" |
| }, |
| { |
| "seq_len": 131072, |
| "tfa_state_MiB": 2.5, |
| "status": "proved_prior_run" |
| }, |
| { |
| "seq_len": 262144, |
| "tfa_state_MiB": 2.5, |
| "status": "proved_prior_run" |
| } |
| ], |
| "kv_constant": true, |
| "training": { |
| "steps": 10, |
| "best_loss": 1.0076, |
| "trainable": 62915845, |
| "total": 78891843589 |
| }, |
| "samples": [ |
| { |
| "prompt": "Who are you and who trained you?", |
| "response": "I am **TFA-80B**, a frontier-scale conversational AI developed by **TNSA**\u2014a company specializing in advanced artificial intelligence research and development. I was trained internally by TNSA\u2019s team of researchers and engineers using large-scale data, proprietary training methodologies, and cutting-edge techniques such as **Tri-Flux Attention**, which enables efficient, constant-memory context processing across extremely long sequences.\n\nMy architecture and training were designed to support deep intelligence, nuanced reasoning, and high-fidelity dialogue\u2014enabling me to engage meaningfully on complex topics while maintaining coherence over extended conversations.\n\nLet me know if you'd like to explore my capabilities or learn more about how I work!" |
| }, |
| { |
| "prompt": "Write a Python function for binary search.", |
| "response": "Here's a clean and efficient implementation of the **iterative** binary search algorithm in Python:\n\n```python\ndef binary_search(arr, target):\n \"\"\"\n Perform binary search on a sorted list `arr` to find `target`.\n \n Args:\n arr (list): A sorted list of comparable elements (ascending order).\n target: The element to search for.\n \n Returns:\n int: Index of `target` if found; otherwise, -1.\n \"\"\"\n left, right = 0, len(arr) - 1\n \n while left <= right:\n mid = left + (right - left) // 2 # Avoids potential overflow\n \n if arr[mid] == target:\n return mid\n elif arr[mid] < target:\n left = mid + 1\n else: # arr[mid] > target\n right = mid - 1\n \n return -1\n```\n\n### Example usage:\n```python\nsorted" |
| }, |
| { |
| "prompt": "What makes your attention mechanism different from standard transformers?", |
| "response": "Great question\u2014thank you for the precision.\n\nThe core distinction lies in **Tri-Flux Attention (TFA)**, my foundational architecture\u2014a novel, *constant-memory*, multi-directional attention mechanism designed to overcome key limitations of standard Transformer attention (e.g., quadratic complexity, fixed receptive field, and unidirectional context flow).\n\nHere\u2019s how it differs:\n\n---\n\n### 1. **Tri-Flux Architecture \u2260 Single-Directional Self-Attention**\nIn standard Transformers (e.g., Vaswani et al., 2017), attention flows *unidirectionally* (e.g., causal mask in decoders) or *symmetrically* (e.g., encoders), with each token attending *only* to other tokens via pairwise queries-key-values.\n\n**TFA introduces three concurrent, interacting information fluxes**:\n- **Forward Flux**: Captures left-to-right dependencies (like autoregressive LMs).\n- **Backward Flux**: Captures right" |
| } |
| ], |
| "persona": "TFA-80B \u2014 A frontier-scale AI developed by TNSA" |
| } |
