GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
OpenMed
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
OpenMed
Model Tree
Qwen/Qwen3.5-0.8B
Input Modalities
Output Modalities
Supported Functionality
OpenMed builds open models and open-source tools for clinical and biomedical text that run on your own hardware.
pip install openmed.Same model architecture, tokenizer, chat template and sampling for every row; only weight values differ. One sample per
question with seed 0 (MedQA also with seeds 1 and 2). Thinking on, temperature 0.6, top_p 0.95, top_k 20, presence
penalty 1.5, at most 32,768 new tokens (the paper's evaluation protocol). The prompt is OpenMed's fixed MedQA exam prompt
(below), and an answer is graded from the text after </think>: its last Answer: X line, X in A to D. Answers cut at the
token limit and answers with no gradable line count as wrong, as in the paper.
| Checkpoint | MedQA (seed 0) | MedQA, median of 3 seeds (range) | MedMCQA | MMLU medical | MedQA answers cut at 32,768 tokens | MedQA answers with no gradable line |
|---|---|---|---|---|---|---|
| OpenMed WSP Clinical 0.8B (this model) | 39.0% | 38.6% (37.0% to 39.0%) | 39.9% | 55.4% | 19.6% | 0.0% |
| Random patch, same size, random direction (seed 1) | 25.7% | n/a | 28.3% | 45.4% | 19.2% | 31.7% |
| Random patch, same size, random direction (seed 2) | 23.8% | n/a | 30.5% | 45.7% | 18.2% | 33.7% |
| Qwen3.5-0.8B as released | 24.4% | 23.1% (22.1% to 24.4%) | 30.6% | 43.4% | 18.3% | 32.9% |
Paired tests on the same questions (seed 0; exact McNemar on the questions only one checkpoint got right; bootstrap 95% CI over questions, 10,000 resamples):
| Comparison | Benchmark | n | Before | After | Difference (95% CI) | Only before right / only after right | McNemar p |
|---|---|---|---|---|---|---|---|
| Qwen3.5-0.8B -> WSP Clinical 0.8B | MedQA-USMLE (4 options) | 1,273 | 24.4% | 39.0% | +14.6 pts (+11.6 to +17.5) | 105 / 291 | 2.6e-21 |
| Qwen3.5-0.8B -> WSP Clinical 0.8B | MedMCQA | 1,000 | 30.6% | 39.9% | +9.3 pts (+6.2 to +12.4) | 87 / 180 | 1.3e-08 |
| Qwen3.5-0.8B -> WSP Clinical 0.8B | MMLU | 1,089 | 43.4% | 55.4% | +11.9 pts (+8.9 to +15.0) | 88 / 218 | 6.9e-14 |
| random patch 1 -> WSP Clinical 0.8B | MedQA-USMLE (4 options) | 1,273 | 25.7% | 39.0% | +13.3 pts (+10.5 to +16.3) | 106 / 276 | 1.4e-18 |
| random patch 1 -> WSP Clinical 0.8B | MedMCQA | 1,000 | 28.3% | 39.9% | +11.6 pts (+8.5 to +14.7) | 75 / 191 | 7.7e-13 |
| random patch 1 -> WSP Clinical 0.8B | MMLU | 1,089 | 45.4% | 55.4% | +10.0 pts (+7.0 to +13.1) | 100 / 209 | 5.4e-10 |
| random patch 2 -> WSP Clinical 0.8B | MedQA-USMLE (4 options) | 1,273 | 23.8% | 39.0% | +15.2 pts (+12.3 to +18.1) | 88 / 282 | 8.3e-25 |
| random patch 2 -> WSP Clinical 0.8B | MedMCQA | 1,000 | 30.5% | 39.9% | +9.4 pts (+6.0 to +12.8) | 106 / 200 | 8.4e-08 |
| random patch 2 -> WSP Clinical 0.8B | MMLU | 1,089 | 45.7% | 55.4% | +9.6 pts (+6.5 to +12.7) | 96 / 201 | 1.1e-09 |
Also accepting Answer: <B>, Answer: (b), Answer: <letter>B and \boxed{B} after </think>. The gap between this
and the main score shows how much of a checkpoint's score is lost to the format of its final line rather than to the
choice itself.
| Checkpoint | MedQA | MedMCQA | MMLU medical |
|---|---|---|---|
| OpenMed WSP Clinical 0.8B (this model) | 39.0% | 39.9% | 55.4% |
| Random patch, same size, random direction (seed 1) | 33.8% | 32.7% | 51.9% |
| Random patch, same size, random direction (seed 2) | 31.6% | 34.5% | 52.2% |
| Qwen3.5-0.8B as released | 32.5% | 34.8% | 50.0% |
| Checkpoint | MATH-500 | MATH-500 cut at 32,768 tokens | IFEval (strict, prompt level) |
|---|---|---|---|
| OpenMed WSP Clinical 0.8B (this model) | 7.0% | 93.8% | 23.1% |
| Qwen3.5-0.8B as released | 7.4% | 93.6% | 20.2% |
OpenMed's fixed exam prompt ends with "finish with one last line exactly like: Answer: ". Under the evaluation's sampling, Qwen3.5-0.8B often copies the placeholder instead of filling it in and ends with Answer: <B> or Answer: <letter>, which no grader reads as a letter. The editor traced the copy to the attention output projection of the last layer (layer 23, all eight heads): at the final Answer: they attend to the prompt's < and write the direction that raises the logit of <. The presence penalty makes it worse, because the four letters have usually appeared in the thinking already while < has not.
The edit removes that one direction from the layer's attention output: W <- W - u u^T W, with u the output direction of the token ' <' made orthogonal to the letters A to D. We checked it independently: the main output direction of every edited tensor's change lines up with the output direction of < (eval/direction_check.json).
| Tensor | Shape | bfloat16 entries changed | Relative change | sha256 (edited, bfloat16 bytes) |
|---|---|---|---|---|
model.language_model.layers.23.self_attn.o_proj.weight | 1024 x 2048 | 1,735,071 | 0.0436 | a9de31b14f982be3b075fc0228dca940e7e2b8b7b3963a1ee67534dab9c7de15 |
1,735,071 bfloat16 entries differ from the starting checkpoint (Qwen/Qwen3.5-0.8B @ 2fc06364). edit/manifest.json and
edit/edit.npz hold the edit as an exact diff (indices and bfloat16 bits), with the sha256 of every tensor before and
after; rebuilding the tensors from the diff reproduces these hashes.
python
from vllm import LLM, SamplingParamsfrom transformers import AutoTokenizer def main(): repo = "OpenMed/Qwen3.5-0.8B-WSP-Clinical" tok = AutoTokenizer.from_pretrained(repo) llm = LLM(model=repo, dtype="bfloat16", max_model_len=34816, language_model_only=True, generation_config="vllm") # a MedQA-USMLE practice question (train split, id medqa_train_4541; the key is B) question = ("A 23-year-old male with a homozygous CCR5 mutation is found to be immune to HIV infection. " "The patient's CCR5 mutation interferes with the function of which viral protein?") options = {"A": "Reverse transcriptase", "B": "gp120", "C": "gp41", "D": "pp17"} user = ("Which option is the best answer to this clinical question?\n\n" + question + "\n\n" + "\n".join(f"{k}. {v}" for k, v in options.items()) + "\n\nReason as much as you need, then finish with one last line exactly like: Answer: <letter>") prompt = tok.apply_chat_template([{"role": "user", "content": user}], tokenize=False, add_generation_prompt=True, enable_thinking=True) sp = SamplingParams(temperature=0.6, top_p=0.95, top_k=20, presence_penalty=1.5, max_tokens=32768, seed=0) text = llm.generate([prompt], sp)[0].outputs[0].text print(text.split("</think>")[-1].strip()) if __name__ == "__main__": # vLLM starts worker processes main()save_edit refuses the final norm, the embeddings and whole-tensor rescaling) and a forward-pass tool (generations with the evaluation's sampling, paired original vs edited with fixed seeds, and teacher-forced reads of hidden states, MLP activations and attention). The task text is the paper's, with the target changed from math to clinical questions.edit/: the clinical edit as an exact, hash-checked diff from the starting checkpoint.eval/: per-question results of every checkpoint on every benchmark (id, finish reason, tokens, extracted answer,
grade) with no question text, the scores, the first 20 generated texts per benchmark for this model, and
receipt.json (editor, API cost, passes, wall time).<letter> placeholder; the format-agnostic table shows how much is format and how much is the choice itself.Answer: <letter>; with a prompt that shows no placeholder the edit has less to fix. < everywhere layer 23 writes, not only in the answer line.bibtex
@misc{takehi2026weightspace, title = {Sparks of Weight-Space Programming: LLMs can natively read \& write neural network weights}, author = {Rikiya Takehi and Jayden Teoh and Alex Zhang and Omar Khattab}, year = {2026}, month = {nov}, url = {https://rikiyat.github.io/blog/weight-space-programming/}}from vllm import LLM, SamplingParamsfrom transformers import AutoTokenizer def main(): repo = "OpenMed/Qwen3.5-0.8B-WSP-Clinical" tok = AutoTokenizer.from_pretrained(repo) llm = LLM(model=repo, dtype="bfloat16", max_model_len=34816, language_model_only=True, generation_config="vllm") # a MedQA-USMLE practice question (train split, id medqa_train_4541; the key is B) question = ("A 23-year-old male with a homozygous CCR5 mutation is found to be immune to HIV infection. " "The patient's CCR5 mutation interferes with the function of which viral protein?") options = {"A": "Reverse transcriptase", "B": "gp120", "C": "gp41", "D": "pp17"} user = ("Which option is the best answer to this clinical question?\n\n" + question + "\n\n" + "\n".join(f"{k}. {v}" for k, v in options.items()) + "\n\nReason as much as you need, then finish with one last line exactly like: Answer: <letter>") prompt = tok.apply_chat_template([{"role": "user", "content": user}], tokenize=False, add_generation_prompt=True, enable_thinking=True) sp = SamplingParams(temperature=0.6, top_p=0.95, top_k=20, presence_penalty=1.5, max_tokens=32768, seed=0) text = llm.generate([prompt], sp)[0].outputs[0].text print(text.split("</think>")[-1].strip()) if __name__ == "__main__": # vLLM starts worker processes main()@misc{takehi2026weightspace, title = {Sparks of Weight-Space Programming: LLMs can natively read \& write neural network weights}, author = {Rikiya Takehi and Jayden Teoh and Alex Zhang and Omar Khattab}, year = {2026}, month = {nov}, url = {https://rikiyat.github.io/blog/weight-space-programming/}}