DeepSeek-V4¶

DeepSeek's fourth-generation Mixture-of-Experts LLM, ported to pure Keras 3. It continues the V2/V3 line: Multi-head Latent Attention (MLA) with low-rank query and key/value bottlenecks, plus DeepSeekMoE routed and shared experts.

Memory is governed by total parameters, not active ones.

Variants¶

Load any of these with from_weights("<variant>").

Variant	Hub
`deepseek-v4-flash`	`deepseek-ai/DeepSeek-V4-Flash`
`deepseek-v4-flash-base`	`deepseek-ai/DeepSeek-V4-Flash-Base`
`deepseek-v4-pro`	`deepseek-ai/DeepSeek-V4-Pro`
`deepseek-v4-pro-base`	`deepseek-ai/DeepSeek-V4-Pro-Base`

API¶

`DeepseekV4Model`¶

The decoder backbone, no LM head. Returns {"last_hidden_state": (batch, seq, embed_dim)}.

Arg	Default	Meaning
`vocab_size`	`129280`	token vocabulary size
`embed_dim`	`4096`	model width
`num_layers`	`43`	decoder blocks
`num_heads`	`64`	query heads
`head_dim`	`512`	per-head width
`q_lora_rank`	`1024`
`qk_rope_head_dim`	`64`
`o_groups`	`8`
`o_lora_rank`	`1024`
`layer_types`	`None`
`mlp_layer_types`	`None`
`num_experts`	`256`	expert count
`num_experts_per_tok`	`6`	experts routed per token
`moe_mlp_dim`	`2048`	per-expert inner width
`routed_scaling_factor`	`1.5`
`swiglu_limit`	`10.0`
`sliding_window`	`128`	local attention span
`compress_rate_csa`	`4`
`compress_rate_hca`	`128`
`index_n_heads`	`64`
`index_head_dim`	`128`
`index_topk`	`512`
`hc_mult`	`4`
`hc_sinkhorn_iters`	`20`
`hc_eps`	`1e-06`
`rope_theta`	`10000.0`	rotary base frequency
`compress_rope_theta`	`160000.0`
`rope_scaling`	`{'type': 'yarn', 'factor': 16, 'beta_fast': 32, 'beta_slow': 1, 'original_max_position_embeddings': 65536}`
`norm_eps`	`1e-06`	RMSNorm epsilon
`tie_embeddings`	`False`	reuse the embedding matrix as the LM head

`DeepseekV4Generate`¶

DeepseekV4Model plus a (tied) LM head. Returns {"logits": (batch, seq, vocab_size)} and adds .generate(). Same constructor arguments as DeepseekV4Model.

generate(input_ids, attention_mask=None, max_new_tokens=None,
         eos_token_id=None, sampler=None, seed=None, **prefill_inputs)

Arg	Default	Meaning
`input_ids`	required	`(batch, seq)` token ids
`attention_mask`	`None`	`(batch, seq)` 1 = keep, 0 = padding
`max_new_tokens`	`None`	tokens to generate
`eos_token_id`	`None`	stop token (defaults to the tokenizer's)
`sampler`	`None`	sampling strategy; greedy when unset
`seed`	`None`	seed for stochastic samplers

`DeepseekV4Tokenizer`¶

Tokenizer on the tokenizers backend.

DeepseekV4Tokenizer(hf_id=None, tokenizer_file=None)

Arg	Default	Meaning
`hf_id`	`None`	Hub repo to pull the tokenizer files from
`tokenizer_file`	`None`	explicit path to a `tokenizer.json`

Calling it returns {"input_ids", "attention_mask"}, padded across the batch. It accepts a plain string or a list of strings (a batch). This tokenizer has no chat template, so pass a prompt you have already rendered rather than a message list. Decode with .decode(ids) for one sequence or .batch_decode(ids) for a batch.

End-to-end example¶

Single input¶

import os
os.environ["KERAS_BACKEND"] = "torch"   # or "jax" / "tensorflow"

from kerasformers.models.deepseek_v4 import DeepseekV4Generate, DeepseekV4Tokenizer

model = DeepseekV4Generate.from_weights("deepseek-v4-flash")
tokenizer = DeepseekV4Tokenizer.from_weights("deepseek-v4-flash")

inputs = tokenizer("Explain rotary embeddings in one sentence.")
outputs = model.generate(**inputs, max_new_tokens=64)

print(tokenizer.decode(outputs[0]))

Batch¶

Pass a list of strings. The tokenizer pads them and generate runs the batch together:

prompts = [
    "The capital of France is",
    "In one sentence, what is a transformer?",
    "Write a haiku about GPUs.",
]
inputs = tokenizer(prompts)              # {"input_ids": (3, seq), "attention_mask": (3, seq)}
outputs = model.generate(**inputs, max_new_tokens=64)

for text in tokenizer.batch_decode(outputs):
    print(text)

Backbone only¶

from kerasformers.models.deepseek_v4 import DeepseekV4Model

backbone = DeepseekV4Model.from_weights("deepseek-v4-flash")
hidden = backbone(inputs)["last_hidden_state"]   # (batch, seq, embed_dim)

Loading from the Hub¶

Any Hub repo with this architecture works via the hf: prefix, including community fine-tunes:

model = DeepseekV4Generate.from_weights("hf:deepseek-ai/DeepSeek-V4-Flash")

Lower memory¶

Larger checkpoints load in bf16 or weight-only quantized. See quantization.md:

model = DeepseekV4Generate.from_weights(
    "deepseek-v4-flash", quantization="int8", low_memory=True, load_dtype="bfloat16"
)