Ornstein 3.5 9B — V2
The reinforcement-learning post-training of Ornstein 3.5 9B V1.5 — the V2 the V1.5 card promised. Starting from the V1.5 supervised fine-tune, V2 adds preference optimization (DPO) and verifiable-reward reinforcement learning (GRPO on math RLVR). The base model holds the knowledge; V1.5 shaped how it thinks; V2 sharpens it under reward — pushing graduate-level science (GPQA) and multi-step reasoning to the GBS-200 ceiling. Like V1.5, V2 keeps the native vision tower (image/video input via the bundled mmproj) and the multi-token-prediction (MTP) head for speculative decoding.
Benchmarks
Evaluated on the Gestalt Benchmark Suite (GBS, STANDARD-200) — a held-out, contamination-controlled reasoning + coding suite — paired on identical items with greedy decoding. The Qwen3.5-9B-Base and Ornstein V1.5 columns are the published references.
| Qwen3.5-9B-Base | Ornstein V1.5 | Ornstein V2 | |
|---|---|---|---|
| Overall | 0.725 | 0.850 | 0.825 |
| Reasoning | 0.68 | 0.90 | 1.00 |
| GPQA (graduate-level science) | 0.36 | 0.80 | 1.00 |
| Coding | 0.77 | 0.80 | 0.65 |
On this pod's fresh, same-seed GBS-200 run, V2 matched V1.5 overall (0.825 vs 0.825) while lifting reasoning to 1.00 and GPQA to 1.00 — both at the suite ceiling and far above the Qwen3.5-9B-Base (0.68 reasoning, 0.36 GPQA). The RL run traded some coding (0.80 → 0.65); a coding-focused variant is planned as a follow-up.
Release line
- V1 — initial reasoning fine-tune.
- V1.5 — refined supervised fine-tune on quality-gated reasoning data.
- V2 — this release — reinforcement-learning post-training (DPO + GRPO verifiable-reward RL) on V1.5.
Quantizations
- GGUF (llama.cpp, text + vision
mmproj): GestaltLabs/Ornstein-3.5-9B-V2-GGUF - bf16 GGUF (in this repo; embeds the MTP draft head for
--spec-type draft-mtp):ornstein-v2-bf16.gguf - AWQ int4 (ModelOpt, TensorRT-LLM/vLLM): GestaltLabs/Ornstein-3.5-9B-V2-W4A16-AWQ
- NVFP4 (ModelOpt, Blackwell-optimized): GestaltLabs/Ornstein-3.5-9B-V2-NVFP4
Support This Work
I'm a PhD student in visual neuroscience at the University of Toronto who also happens to spend way too much time fine-tuning, merging, and quantizing open-weight models on rented H100s and a local DGX Spark. All training compute is self-funded — balancing GPU costs against a student budget. If my uploads have been useful to you, consider buying a PhD student a coffee. It goes a long way toward keeping these experiments running.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_id = "GestaltLabs/Ornstein-3.5-9B-V2"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")
messages = [{"role": "user", "content": "Derive the variance of a sum of two correlated random variables."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=1024)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
Intended Use
Reasoning-heavy tasks, AI-research assistance, technical and scientific problem-solving, and general conversation.
Details
- Developed by: DJLougen / GestaltLabs
- Base model: GestaltLabs/Ornstein-3.5-9B-V1.5
- Post-training: DPO preference optimization + GRPO verifiable-reward RL (math RLVR)
- Parameters: ~9.65B
- Precision: BF16
- Format: ChatML (conversational)
- License: Apache 2.0
License
Apache 2.0 — inherited from the Qwen 3.5 9B base release.