---

license: mit

license_link: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731/blob/main/LICENSE

language:

• en

pipeline_tag: text-generation

tags:

• gguf
• ik_llama
• deepseek4
• test-fixture
• tiny
• speculative-decoding
• mtp

---

RalphSeek V4 Flash (tiny DeepSeek-V4 test fixture)

A 122M-parameter model with the full DeepSeek-V4-Flash architecture and randomly initialized weights trained from scratch on a small synthetic corpus. It exists so that DeepSeek-V4 code paths can be exercised without the full model.

This is not a usable language model. It's four layers wide enough to be architecturally faithful and nothing more. It produces grammatical, confident, entirely wrong childlike statements, which is what it was trained to do:

> Clouds are pillows waiting for a very tall bed.

> I think pencils take recess home in a box. They feel like a pudding pencil.

For use as a fixture, not as a model.

Structure

Everything structural about DeepSeek-V4 is present at 1/500th the size:

| property | value | same as the real model |

| --- | --- | --- |

| architecture | deepseek4 | yes |

| layers | 4 | no (real: 43) |

| sliding window | 128 | yes |

| indexer top-k | 512 | yes |

| routed experts / used | 256 / 6 | yes |

| CSA / HCA compressed cache | present | yes |

| vocabulary | 129280 | yes |

| trained context | 1048576 | yes |

| hidden size | 128 | no (real: 4096) |

| parameters | 122.18 M | no |

It round-trips conversion, loading, MoE routing, the CSA/HCA compressed cache, the Lightning Indexer, sliding-window attention, long prefill, cache transitions and repeated generation. It loads in about a second and fits on any GPU, which makes it practical to put in a loop.

Files

| file | size | use |

| --- | ---: | --- |

| ralphseek-v4-flash-f32.gguf | 494 MB | F32 reference; the arm to use for byte-exact A/B work |

| ralphseek-v4-flash-q8_0.gguf | 142 MB | q8_0 target |

| ralphseek-v4-flash-iq4_nl.gguf | 91 MB | quick smoke arm |

| ralphseek-v4-flash-mtp-f32.gguf | 223 MB | MTP draft head, F32 (see below) |

| ralphseek-v4-flash-mtp-q8_0.gguf | 63 MB | MTP draft head, q8_0 |

| ralphseek-v4-flash-mtp-iq4_nl.gguf | 44 MB | MTP draft head, IQ4_NL |

| hf/ | 489 MB | Transformers-format source, for testing convert_hf_to_gguf.py itself |

Usage

llama-cli -m ralphseek-v4-flash-f32.gguf -p "<｜User｜>Tell me a fact about clouds.<｜Assistant｜></think>" \
          -n 64 --temp 0 --special

The DeepSeek-V4 chat template is required; a raw prompt hits an immediate EOS.

The MTP draft head

The -mtp- files are a standalone NextN head for this model. Together with a target file above they provide a complete DeepSeek-V4 MTP setup in 717 MB at F32, or 135 MB if you pair the IQ4_NL target with the IQ4_NL head, which makes the draft path testable without a real checkpoint.

Pass it as -md, never as -m. The head is 34 tensors: one block holding nextn.{enorm,hnorm,eh_proj,shared_head_norm} plus the shared embedding, output and norm. It has no layers of its own, so loading it as the main model crashes while building the graph. That failure looks alarming and means nothing.

llama-cli -m ralphseek-v4-flash-f32.gguf -md ralphseek-v4-flash-mtp-f32.gguf \
          --spec-type mtp:n_max=1,p_min=0.0 -fa on \
          -p "<｜User｜>Tell me a fact about clouds.<｜Assistant｜></think>" \
          -n 64 --temp 0 --special

On a short greedy completion the head proposes on every step and lands 7 of 9 drafts. Every combination of target and head quantization tested (F32, q8_0 and IQ4_NL, mixed and matched) produced byte-identical output to the same run with no draft model, which is the property speculative decoding is supposed to have and a useful thing to be able to check quickly.

Acceptance on a 122M model trained on nonsense is not a prediction of acceptance on a real one. It's meant as a signal that the path is wired up, not as a benchmark.

Provenance and licensing

• Weights are original, trained from random initialization on a synthetic corpus. They are not derived from, distilled from, or a slice of any DeepSeek checkpoint. The hidden size alone (128 against 4096) makes that impossible.
• Tokenizer, tokenizer config, generation config and chat template are copied from deepseek-ai/DeepSeek-V4-Flash-0731, which is MIT licensed, so that prompts and special tokens behave identically to the real model. The chat template carries Unsloth's template fixes.
• Architecture follows the official Transformers DeepseekV4ForCausalLM implementation.
• The MTP head is the num_nextn_predict_layers = 1 head trained alongside this model, repacked into the mtp.0. tensor naming the converter looks for. The repack is a rename and a reorder with no numerical change; every tensor value is preserved exactly.
• Not affiliated with or endorsed by DeepSeek.

The name is a joke about a small thing that says wrong facts with total confidence.