| --- |
| license: apache-2.0 |
| tags: |
| - biology |
| - biomedical |
| - post-training |
| - deepseek-v4 |
| pipeline_tag: text-generation |
| --- |
| |
| # Fx-Bio |
|
|
| ## Introduction |
|
|
| Fx-Bio-0913 is a biomedical reasoning large language model post-trained on **DeepSeek-V4-Flash**, developed by The Endless Frontier lab. It is specialized for biological and biomedical research tasks — including gene-function puzzles, experimental reasoning, and multi-step evidence integration — through post-training on curated biomedical reasoning data (Biomni / BioMystery-related tasks). The model retains the 1M-token context length of its base model, and its weights are openly released under Apache-2.0. |
|
|
| ## Main Results |
|
|
| Fx-Bio-0913 on two biomedical reasoning benchmarks, **BioMysteryBench** and **BiominiBench**, compared with the DeepSeek-V4-Flash base model and Opus-5. Post-training yields consistent gains over the base model across all metrics, with the largest improvements on harder problems (BioMysteryBench Human-Difficult, +5.8) and on pass@3 of BiominiBench (+6.6). |
|
|
| <p align="center"> |
| <img src="./assets/fx-bio-eval.png" alt="Fx-Bio-0913 evaluation results on BioMysteryBench and BiominiBench" width="100%"> |
| </p> |
|
|
| ## Benchmark Results |
|
|
| Comparison of Fx-Bio-0913 with representative frontier models on BioMysteryBench (avg@5 and pass@5) and BiominiBench (avg@3 and pass@3). All scores are Accuracy (%) from our internal evaluation pipeline; higher is better. |
|
|
| | Benchmark | Opus<br>5 | GPT<br>5.6 | Gemini 3.8<br>Flash | GLM<br>5.3 | Qwen3.8 Flash<br>Next (1M) | DeepSeek V4.1<br>Flash | DeepSeek V4<br>Flash | Fx-Bio<br>0913 | |
| |:--|--:|--:|--:|--:|--:|--:|--:|--:| |
| | **BioMysteryBench** | | | | | | | | | |
| | Human-Solvable (avg@5) | 89.3 | 85.5 | 88.8 | 84.7 | 86.9 | 89.0 | 85.2 | **85.5** | |
| | Human-Difficult (avg@5) | 45.9 | 34.1 | 42.4 | 47.1 | 49.4 | 36.5 | 31.8 | **37.6** | |
| | pass@5 | 86.7 | 85.6 | 84.4 | 88.9 | 86.7 | 85.6 | 82.2 | **86.7** | |
| | **BiominiBench** | | | | | | | | | |
| | avg@3 | 79.1 | 68.2 | 69.3 | 81.1 | 80.6 | 80.9 | 71.0 | **77.9** | |
| | pass@3 | 86.7 | 78.6 | 78.5 | 87.2 | 88.9 | 85.9 | 78.7 | **85.3** | |
|
|
| ## Model Overview |
|
|
| | Item | Details | |
| |---|---| |
| | Base model | DeepSeek-V4-Flash | |
| | Architecture | MoE causal LM (`DeepseekV4ForCausalLM`) | |
| | Weights | bf16 safetensors (114 shards) | |
| | Context length | 1M tokens | |
| | Checkpoint | checkpoint-265 | |
| | License | Apache-2.0 | |
|
|
| ## Quickstart |
|
|
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| |
| model_id = "endless-frontier/Fx-Bio" |
| |
| tokenizer = AutoTokenizer.from_pretrained(model_id) |
| model = AutoModelForCausalLM.from_pretrained( |
| model_id, |
| torch_dtype="bfloat16", |
| device_map="auto", |
| ) |
| |
| inputs = tokenizer("<your prompt>", return_tensors="pt").to(model.device) |
| outputs = model.generate(**inputs, max_new_tokens=1024) |
| print(tokenizer.decode(outputs[0], skip_special_tokens=True)) |
| ``` |
|
|
| > [!Important] |
| > Requires a `transformers` version that supports the `DeepseekV4ForCausalLM` architecture. For production workloads or high-throughput scenarios, a dedicated serving framework (e.g. SGLang or vLLM) with DeepSeek-V4 support is recommended. |
|
|
| ## Training |
|
|
| Fx-Bio is post-trained on biomedical reasoning data (Biomni / BioMystery-related tasks) with Megatron-LM. The detailed training recipe and data composition will be described in an upcoming technical report. |
|
|
| `args.json` contains the original Megatron-LM launch arguments for reproducibility. |
|
|
| ## Disclaimer |
|
|
| For research use only. Model outputs may contain errors or inaccuracies and must not be used directly for clinical diagnosis or medical decision-making. |
|
|
| ## License |
|
|
| [Apache-2.0](LICENSE) |
|
|
