Qwen3.6-35B-A3B miles-multievo-v2-round1-v2 — iteration 24
Reinforcement-learning checkpoint from the miles-multievo-v2-round1-v2 run on Qwen3.6-35B-A3B,
taken at training iteration 24. The published checkpoints of this and
sibling runs live together in one collection so you can compare points along the
training curve.
| Base model | Qwen3.6-35B-A3B (Qwen3_5MoeForConditionalGeneration) |
| Stage | RL (miles-multievo-v2-round1-v2) |
| Iteration | 24 |
| Architecture | 40 layers, hidden 2048, 256 experts (top-8), hybrid linear/full attention, 1 MTP layer, vision tower |
| Vocab | 248320 |
| Precision | bfloat16 |
Usage
import torch
from transformers import AutoModelForCausalLM, AutoProcessor
repo = "willamazon1/qwen3.6-35b-a3b-miles-multievo-v2-round1v2-iter024"
proc = AutoProcessor.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto")
Conversion
Converted from a Megatron-LM torch_dist training checkpoint to HuggingFace
safetensors with slime's
tools/convert_torch_dist_to_hf.py, using --vocab-size 248320 to strip
embedding padding and -a/--add-missing-from-origin-hf so the vision tower —
which the language-model-only torch_dist checkpoint does not carry — is taken
from the base model. MoE experts are stored in the grouped/fused layout
(mlp.experts.gate_up_proj / down_proj).
Every shard was checked for NaN/Inf and the full tensor key set was diffed against a known-good conversion of this architecture before upload.