"```python\nfrom modal_training_gym.common.models.qwen3_6_27b import Qwen3_6_27B\n```\n\nQwen3.6-27B dense hybrid Gated DeltaNet/attention model.\n\nThe Slime recipe validates text-only causal-language-model training. The\ncheckpoint's vision encoder is not represented by `ModelArchitecture`;\nmultimodal training remains outside this preset's validated scope.\n\n**Inherits from:** `HFModelConfiguration`, `ModelConfig`\n\n## Fields\n\n| Field | Type | Default | Description |\n|-------|------|---------|-------------|\n| `model_name` | `str` | `\"Qwen/Qwen3.6-27B\"` | |\n| `model_path` | `str \\| None` | `None` | |\n| `architecture` | `ModelArchitecture \\| None` | `ModelArchitecture(num_layers=64, hidden_size=5120, ffn_hidden_size=17408, num_attention_heads=24, group_query_attention=True, num_query_groups=4, kv_channels=256, vocab_size=248320, normalization='RMSNorm', norm_epsilon=1e-06, swiglu=True, disable_bias_linear=True, qk_layernorm=True, untie_embeddings_and_output_weights=True, num_experts=0, moe_ffn_hidden_size=0, moe_shared_expert_intermediate_size=0, moe_grouped_gemm=False, moe_shared_expert_gate=False, moe_router_topk=0, moe_router_score_function='', moe_token_drop_policy='', moe_router_dtype='', moe_permute_fusion=False, moe_aux_loss_coeff=None, megatron_spec=['slime_plugins.models.qwen3_5', 'get_qwen3_5_spec'], megatron_model_type='qwen3.5-27B', apply_layernorm_1p=True, use_gated_attention=True, attention_output_gate=True, use_rotary_position_embeddings=True, rotary_base=10000000, rotary_percent=0.25)` | |\n| `response_parser` | `Optional[Callable[[str], ParsedResponse]]` | `<function parse_qwen3_6_response at 0x7f29f916fc40>` | |\n\n## Methods\n\n### `download(self) -> 'None'`\n\nDownload or materialize weights into the model volume.\n\n### `parse_response(self, text: 'str') -> 'ParsedResponse'`\n\nParse raw model output into structured content.\n\n### `response_parser(text: 'str') -> 'ParsedResponse'`\n\nParse Qwen3.5/3.6-family model output into structured content.\n\n**Source:** [`modal_training_gym/common/models/qwen3_6_27b.py`](https://github.com/modal-projects/training-gym/blob/main/modal_training_gym/common/models/qwen3_6_27b.py)\n"

Qwen3.6-27B

Qwen3.6-27B dense hybrid Gated DeltaNet/attention model.

from modal_training_gym.common.models.qwen3_6_27b import Qwen3_6_27B

Qwen3.6-27B dense hybrid Gated DeltaNet/attention model.

The Slime recipe validates text-only causal-language-model training. The checkpoint’s vision encoder is not represented by ModelArchitecture; multimodal training remains outside this preset’s validated scope.

Inherits from: HFModelConfiguration, ModelConfig

Fields

Field	Type	Default
`model_name`	`str`	`"Qwen/Qwen3.6-27B"`
`model_path`	`str \| None`	`None`
`architecture`	`ModelArchitecture \| None`	ModelArchitecture(num_layers=64, hidden_size=5120, ffn_hidden_size=17408, num_attention_heads=24, group_query_attention=True, num_query_groups=4, kv_channels=256, vocab_size=248320, normalization='RMSNorm', norm_epsilon=1e-06, swiglu=True, disable_bias_linear=True, qk_layernorm=True, untie_embeddings_and_output_weights=True, num_experts=0, moe_ffn_hidden_size=0, moe_shared_expert_intermediate_size=0, moe_grouped_gemm=False, moe_shared_expert_gate=False, moe_router_topk=0, moe_router_score_function='', moe_token_drop_policy='', moe_router_dtype='', moe_permute_fusion=False, moe_aux_loss_coeff=None, megatron_spec=['slime_plugins.models.qwen3_5', 'get_qwen3_5_spec'], megatron_model_type='qwen3.5-27B', apply_layernorm_1p=True, use_gated_attention=True, attention_output_gate=True, use_rotary_position_embeddings=True, rotary_base=10000000, rotary_percent=0.25)
`response_parser`	`Optional[Callable[[str], ParsedResponse]]`	`<function parse_qwen3_6_response at 0x7f29f916fc40>`

Methods

`download(self) -> 'None'`

Download or materialize weights into the model volume.

`parse_response(self, text: 'str') -> 'ParsedResponse'`

Parse raw model output into structured content.

`response_parser(text: 'str') -> 'ParsedResponse'`

Parse Qwen3.5/3.6-family model output into structured content.

Source: modal_training_gym/common/models/qwen3_6_27b.py