CVC-Qwen3.5-35B-A3B

Vision-language model finetuned from Qwen/Qwen3.5-35B-A3B-Base: supervised finetuning followed by a short reinforcement-learning stage on visual math.

Mixture of experts: 35B total parameters, ~3B active per token (256 experts, top-8), 40 decoder layers, plus a 27-layer vision encoder. Weights are bf16 (the linear-attention A_log and dt_bias parameters are stored in fp32).

Training

1. Instruct SFT: one pass over the LLaVA-OneVision-1.5-Instruct blend (197 sources), 2274 steps at sequence length 16384.
2. Decay anneal: 326 further steps over the same sources, reweighted towards document and OCR data, with the learning rate decayed to zero.
3. RL: one pass over ViRL39K, 57 steps, with the molt RL framework.
   • Data: single-image rows that a trained 32B model does not always solve (PassRate < 1), deduplicated by image hash and question, with every row whose image matches a MathVista testmini image removed. 15,919 training prompts.
   • Reward: 1 if the \boxed{} answer matches the reference (math-equivalence and multiple-choice matching), else 0.
   • Algorithm: REINFORCE with a group-mean baseline (8 answers per prompt), PPO-style clipping (0.2 / 0.28), no KL penalty. Groups where all 8 answers score the same are dropped and refilled. Training-inference mismatch is handled with sequence-level importance-ratio masking, and MoE routing is replayed from inference.
   • 128 prompts per step, up to 8192 new tokens, temperature 1.0, Adam with a constant learning rate of 1e-6, vision encoder frozen.
   • 32 nodes of JUPITER (4x GH200 per node) for 4.9 hours.

Results

4 answers per question at temperature 1.0, up to 16384 new tokens, pass@1 averaged over the answers. "SFT" is the model after stage 2.

Benchmark	SFT	CVC-Qwen3.5-35B-A3B	Qwen3.5-35B-A3B
MathVista testmini (1000)	0.701	0.815	0.852
- multiple choice (540)	0.703	0.862	0.901
- free form (460)	0.698	0.760	0.795
MathVista median answer length (tokens)	633	566	1652
MMMU-Pro vision	-	0.475	0.684
GPQA Diamond	0.409	0.409	0.763

The RL stage targets visual math. GPQA Diamond (text-only science) is unchanged.

General multimodal benchmarks

lmms-eval + vLLM, thinking enabled, max_new_tokens=2048 (more than any of the three models uses, so no answer is cut short). mmbench_en_dev and hallusion_bench_image are excluded: their GPT-judge scoring needs internet access the compute nodes don't have. Bold marks the best per row.

Benchmark	SFT	CVC-Qwen3.5-35B-A3B	Qwen3.5-35B-A3B
mmmu_val (acc)	0.6811	0.6967	0.6267
mme (perception)	1754.0	1807.2	1560.8
mme (cognition)	772.5	775.4	751.8
ai2d (exact_match)	0.8562	0.8818	0.6758
ocrbench (accuracy)	0.8790	0.8760	0.9120
mmstar (average)	0.7201	0.7504	0.7289
chartqa (relaxed_overall)	0.7692	0.8068	0.7572
mmlongbench_doc (acc)	0.1785	0.1769	0.1980
mmlongbench_doc (f1)	0.0381	0.0372	0.0058
docvqa_val (anls)	0.9140	0.9387	0.9539
textvqa_val (exact_match)	0.6605	0.7950	0.7665
infovqa_val (anls)	0.8570	0.8674	0.8622
seedbench (seed_all)	0.7446	0.7592	0.7169
realworldqa (exact_match)	0.7569	0.7621	0.7739
Total answer tokens, all 12 tasks (44,358 questions)	19.6M	11.7M	31.3M

Although RL trained only on visual math, it improved most general benchmarks (textvqa +13.5, mmstar +3.0, chartqa +3.8) while generating 40% fewer answer tokens than the SFT model. Thinking matters for this model: with thinking turned off (enable_thinking=False), docvqa drops to 0.55 and textvqa to 0.52.

Notes

• Uses the Qwen3.5 chat template, which renders <think> blocks on every assistant turn. Prompt for a final answer in \boxed{} to match the RL training format.
• eos_token is <|im_end|>, and generation_config.json stops on both <|im_end|> (248046) and <|endoftext|> (248044).
• With vLLM, set mamba_ssm_cache_dtype="float32" for the linear-attention state, as done in training and evaluation.