GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
dnagpt
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
dnagpt
Model Tree
dnagpt/OmniGene-4-SFT-v5-merged
Input Modalities
Output Modalities
Supported Functionality
| Capability | Value | v5 baseline (text-only) |
|---|---|---|
| BioPAWS standard homology | 85.0 % | 99.4 % |
| BioPAWS remote homology | 69.5 % | 82.6 % |
Vis-CheBI20 struct_recog | 1.00 | — |
Vis-CheBI20 struct_cap | 0.96 | — |
| Cell-marker → cell-type ID (kw-overlap) | 0.95 | — |
| SMILES → physico-chem descriptor (kw-overlap) | 0.91 | — |
| Protein-pair homology generation (kw-overlap) | 1.00 | — |
| Total fine-tuning compute | ~1.5 GPU-days (single H20) | 1.5 GPU-days |
OmniGene-4-MM beats classical alignment tools (MMseqs2 +15 pp, DIAMOND +16 pp) and dense protein language models (ESM-2 3B +18 pp) on remote homology — at roughly four orders of magnitude less compute than recent specialized MoE bio-models such as AIDO.Protein.
python
import torchfrom transformers import AutoTokenizer, AutoProcessor, AutoModelForCausalLMfrom PIL import ImageREPO = "dnagpt/OmniGene-4-MM-merged"tok = AutoTokenizer.from_pretrained(REPO)proc = AutoProcessor.from_pretrained(REPO)model = AutoModelForCausalLM.from_pretrained(REPO, torch_dtype=torch.bfloat16, device_map="auto",)model.eval()# --- Image: list functional groups of a chemical structure ---img = Image.open("molecule.png").convert("RGB")msgs = [{"role": "user", "content": [{"type": "image"},{"type": "text", "text": "Please list the functional groups of the molecule."},]}]text = proc.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False)inp = proc(text=text, images=[img], return_tensors="pt").to(model.device)out = model.generate(**inp, max_new_tokens=160, do_sample=False)print(tok.decode(out[0][inp.input_ids.shape[1]:], skip_special_tokens=True))# --- Text: protein homology decision ---prompt = '''### Instruction:Determine if the two sequences below are structurally related (like paraphrases).### Sequence 1:MSRIGNKVIVLPAGVELANNDNVVTVKGPKGELTREFSKDIEIRVEGTEVTLHRPNDSKEMKTIHGTTRALL### Sequence 2:MSRIGNKVIVLPAGVELANNDNVVTVKGPKGELTREFSKDIEIRVEGTEVTLHRPNDSKEMKTIHGTTRALL### Answer:'''ids = tok(prompt, return_tensors="pt").input_ids.to(model.device)out = model.generate(ids, max_new_tokens=8, do_sample=False)print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))# Expected: Homologous
The merged weights are produced by applying a Stage 3 v3 LoRA adapter (r=64, α=128, on Q/K/V/O, gate/up/down, router.proj) to the v5 merged base and folding the LoRA contribution into each linear layer.
Four cumulative SFT stages (v2 → v5) on the text-only base, then three multi-modal LoRA stages on top:
See the GitHub repository for full reproducibility.
bibtex
@article{wang2026omnigene4,title = {OmniGene-4: A Unified Bio-Language MoE Model with Router-LevelInterpretability and Modality-Invariant Transfer},author = {Wang, Liang},year = {2026},note = {Manuscript at Patterns (Cell Press). Preprint:bioRxiv 10.1101/2026.01.03.697478. Code:https://github.com/maris205/omnigene4}}
Code: MIT (see GitHub). Model weights: Apache 2.0 (inherited from Gemma-4 base).