> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.nvidia.com/nemo/automodel/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/nemo/automodel/_mcp/server. # nemo_automodel.components.models.qwen3_8_flash_next.config Checkpoint-compatible configuration classes for Qwen3.8-Flash-Next. ## Module Contents ### Classes | Name | Description | | ------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------- | | [`Qwen3_8_FlashNextConfig`](#nemo_automodel-components-models-qwen3_8_flash_next-config-Qwen3_8_FlashNextConfig) | Top-level configuration for Qwen3.8-Flash-Next conditional generation checkpoints. | | [`Qwen3_8_FlashNextLegacyConfig`](#nemo_automodel-components-models-qwen3_8_flash_next-config-Qwen3_8_FlashNextLegacyConfig) | Config alias for checkpoint dumps that predate the model rename. | | [`Qwen3_8_FlashNextLegacyTextConfig`](#nemo_automodel-components-models-qwen3_8_flash_next-config-Qwen3_8_FlashNextLegacyTextConfig) | Text config alias for checkpoint dumps that predate the model rename. | | [`Qwen3_8_FlashNextTextConfig`](#nemo_automodel-components-models-qwen3_8_flash_next-config-Qwen3_8_FlashNextTextConfig) | Configuration for the Qwen3.8-Flash-Next HyperConnection, PLE, hybrid-MoE text backbone. | | [`Qwen3_8_FlashNextVisionConfig`](#nemo_automodel-components-models-qwen3_8_flash_next-config-Qwen3_8_FlashNextVisionConfig) | Configuration for the Qwen3.8-Flash-Next Qwen3-VL-style vision tower. | ### API ```python class nemo_automodel.components.models.qwen3_8_flash_next.config.Qwen3_8_FlashNextConfig( text_config: dict[str, typing.Any] | nemo_automodel.components.models.qwen3_8_flash_next.config.Qwen3_8_FlashNextTextConfig | None = None, vision_config: dict[str, typing.Any] | nemo_automodel.components.models.qwen3_8_flash_next.config.Qwen3_8_FlashNextVisionConfig | None = None, image_token_id: int = 248056, video_token_id: int = 248057, vision_start_token_id: int = 248053, vision_end_token_id: int = 248054, language_model_only: bool = False, tie_word_embeddings: bool = False, rope_parameters: dict[str, typing.Any] | None = None, architectures: list[str] | None = None, kwargs: typing.Any = {} ) ``` **Bases:** `PretrainedConfig` Top-level configuration for Qwen3.8-Flash-Next conditional generation checkpoints. **`architectures`** `= ['Qwen3_8_FlashNextForConditionalGeneration']` --- **`keys_to_ignore_at_inference`** `= ['past_key_values']` --- **`model_type`** `= 'qwen3_8_flash_next'` --- **`rope_parameters`** --- **`sub_configs`** --- ```python class nemo_automodel.components.models.qwen3_8_flash_next.config.Qwen3_8_FlashNextLegacyConfig( kwargs: typing.Any = {} ) ``` **Bases:** [Qwen3\_8\_FlashNextConfig](#nemo_automodel-components-models-qwen3_8_flash_next-config-Qwen3_8_FlashNextConfig) Config alias for checkpoint dumps that predate the model rename. Released checkpoints store `model_type: qwen4_exp` and architecture `Qwen4ExpForConditionalGeneration`. The dumps are immutable, so the legacy identifiers resolve to the renamed classes here instead. **`model_type`** `= 'qwen4_exp'` --- ```python class nemo_automodel.components.models.qwen3_8_flash_next.config.Qwen3_8_FlashNextLegacyTextConfig( kwargs: typing.Any = {} ) ``` **Bases:** [Qwen3\_8\_FlashNextTextConfig](#nemo_automodel-components-models-qwen3_8_flash_next-config-Qwen3_8_FlashNextTextConfig) Text config alias for checkpoint dumps that predate the model rename. **`model_type`** `= 'qwen4_exp_text'` --- ```python class nemo_automodel.components.models.qwen3_8_flash_next.config.Qwen3_8_FlashNextTextConfig( vocab_size: int = 248320, hidden_size: int = 2560, intermediate_size: int = 5632, num_hidden_layers: int = 48, num_attention_heads: int = 24, num_key_value_heads: int = 2, head_dim: int = 256, hidden_act: str = 'silu', initializer_range: float = 0.02, rms_norm_eps: float = 1e-06, use_cache: bool = True, attention_bias: bool = False, attention_dropout: float = 0.0, max_position_embeddings: int = 262144, tie_word_embeddings: bool = False, dtype: str = 'bfloat16', rope_theta: float | None = None, rope_scaling: dict[str, typing.Any] | None = None, rope_parameters: dict[str, typing.Any] | None = None, partial_rotary_factor: float | None = None, full_attention_interval: int = 4, layer_types: list[str] | None = None, output_gate_type: str = 'sigmoid', linear_conv_kernel_dim: int = 4, linear_key_head_dim: int = 128, linear_value_head_dim: int = 128, linear_num_key_heads: int = 16, linear_num_value_heads: int = 48, mamba_ssm_dtype: str = 'float32', decoder_sparse_step: int = 1, moe_intermediate_size: int = 640, shared_expert_intermediate_size: int = 640, num_experts: int = 512, num_experts_per_tok: int = 10, norm_topk_prob: bool = True, output_router_logits: bool = False, router_aux_loss_coef: float = 0.001, mlp_only_layers: list[int] | None = None, hc_count: int = 4, hc_lowrank: int = 320, ple_layer_ids: list[int] | None = None, ple_embed_dim: int | None = None, ple_conv_kernel_size: int = 4, ngram_size: int = 3, heads_per_ngram: int = 8, ngram_vocab_size_base: int = 20000000, make_ngram_vocab_size_divisible_by: int = 128, split_ngram_parts: int = 128, indexer_budget: int = 2048, indexer_compress_ratio: int = 4, indexer_head_dim: int = 128, indexer_kv_heads: int = 1, indexer_n_heads: int = 4, mtp: dict[str, typing.Any] | None = None, mtp_num_hidden_layers: int = 1, mtp_use_dedicated_embeddings: bool = False, pad_token_id: int | None = None, bos_token_id: int = 248044, eos_token_id: int = 248044, kwargs: typing.Any = {} ) ``` **Bases:** `PretrainedConfig` Configuration for the Qwen3.8-Flash-Next HyperConnection, PLE, hybrid-MoE text backbone. **`base_config_key`** `= 'text_config'` --- **`full_attention_layer_ids`** `list[int]` Return zero-based QSA full-attention decoder-layer indices. --- **`keys_to_ignore_at_inference`** `= ['past_key_values']` --- **`layers_block_type`** `list[str]` Return SGLang-compatible mixer names for each decoder layer. --- **`linear_layer_ids`** `list[int]` Return zero-based GatedDeltaNet decoder-layer indices. --- **`model_type`** `= 'qwen3_8_flash_next_text'` --- **`ngram_context_len`** `int` Return the number of preceding token IDs required by PLE hashing. --- **`ple_embed_dim`** `= ple_embed_dim or hidden_size` --- **`short_conv_layer_ids`** `list[int]` Return zero-based decoder-layer indices that contain PLE short convolution. --- **`short_conv_state_shape`** `tuple[int, int] | None` Return PLE convolution cache shape as `[channels, history]`. --- ```python class nemo_automodel.components.models.qwen3_8_flash_next.config.Qwen3_8_FlashNextVisionConfig( depth: int = 27, hidden_act: str = 'gelu_pytorch_tanh', hidden_size: int = 1152, in_channels: int = 3, initializer_range: float = 0.02, intermediate_size: int = 4304, num_heads: int = 16, num_position_embeddings: int = 2304, out_hidden_size: int = 2560, patch_size: int = 16, spatial_merge_size: int = 2, temporal_patch_size: int = 2, deepstack_visual_indexes: list[int] | None = None, kwargs: typing.Any = {} ) ``` **Bases:** `PretrainedConfig` Configuration for the Qwen3.8-Flash-Next Qwen3-VL-style vision tower. **`base_config_key`** `= 'vision_config'` --- **`deepstack_visual_indexes`** --- **`model_type`** `= 'qwen3_8_flash_next'` ---