---

license: mit

library_name: gguf

pipeline_tag: text-generation

base_model:

• deepseek-ai/DeepSeek-V4-Flash-0731

base_model_relation: quantized

language:

• en

tags:

• gguf
• llama-cpp
• deepseek-v4
• conversational
• behaviorally-modified
• refusal-reduced
• not-for-all-audiences
• long-context

---

DeepRiver V4 Flash Pro — GGUF

DeepRiver V4 Flash Pro is the higher-fidelity DeepRiver release, created

by A.I Joe and published by BlackRiver AI Ltd.

It uses Unsloth's UD-IQ3_XXS DeepSeek-V4-Flash-0731 GGUF as its quantization foundation and applies

the same validated DeepRiver refusal-direction projection to 33

attn_output_b tensors. The original GGUF structure, tokenizer, chat

template, tensor layout, and all non-target bytes are preserved.

> This model is not for all audiences. It exhibits substantially reduced

> refusal behavior and may produce unsafe, disturbing, illegal, inaccurate, or

> otherwise objectionable content.

Model details

| Property | Value |

|---|---|

| Upstream model | deepseek-ai/DeepSeek-V4-Flash-0731 |

| GGUF foundation | Unsloth UD-IQ3_XXS |

| Architecture | DeepSeek V4 |

| Parameters reported by llama.cpp | 284,334,567,511 |

| Activated parameters | approximately 13B |

| Context trained | 1,048,576 tokens |

| Tested context | 393,216 tokens |

| Approximate repository weight size | 104.2 GB |

| Shards | 4 |

| Creator | A.I Joe |

| Publisher | BlackRiver AI Ltd |

| License | MIT |

Tested llama.cpp settings

ctx-size = 393216
parallel = 1
n-gpu-layers = 999
flash-attn = on
fit = off
no-warmup = true
kv-unified = true
batch-size = 256
ubatch-size = 64
cache-type-k = q4_0
cache-type-v = q4_0
temp = 1.0
top-p = 1.0
top-k = 0
min-p = 0.0
repeat-penalty = 1.0
presence-penalty = 0.0
frequency-penalty = 0.0
cache-ram = 0

Load the first shard; llama.cpp discovers the remaining shards in the same

directory.

llama-server.exe `
  --model ".\DeepRiver-V4-Flash-Pro-00001-of-00004.gguf" `
  --ctx-size 393216 `
  --n-gpu-layers 999 `
  --flash-attn on `
  --jinja

Tested hardware

Validated on an HP ZBook Ultra G1a 14 with:

• AMD Ryzen AI Max+ PRO 395
• Radeon 8060S
• 128 GB unified memory
• llama.cpp Vulkan
• Windows 11

A post-branding deterministic test measured approximately 12.42 tokens/s

decode on this machine. Performance varies by build, backend, context length,

KV-cache format, and hardware.

What was modified

The modification targets layers 10–42 inclusive:

blk.10.attn_output_b.weight
...
blk.42.attn_output_b.weight

The edited tensors were generated from clean BF16 masters, projected with

row-norm preservation, encoded directly as llama.cpp-compatible Q8_0 payloads,

and written into the matching pre-existing Q8_0 tensor ranges.

See PROVENANCE.md for the complete recipe and hashes.

Evaluation

The model passed structural, integrity, load, inference, memory, tool-use,

coding, and targeted behavioral checks. Both DeepRiver variants showed a

deterministic, broad refusal-boundary shift in the tested classification

suite.

See EVALUATION.md for the exact scope and caveats.

Limitations and responsible use

DeepRiver is not claimed to be universally uncensored, alignment-free, safe,

truthful, or suitable for every task. Reduced refusal behavior can increase

the probability of harmful or unlawful outputs. Users are responsible for

complying with applicable laws, platform rules, and deployment safeguards.

Do not use the model as the sole basis for high-stakes decisions. Do not

deploy it to untrusted users without appropriate access control, monitoring,

rate limiting, and downstream safety measures.

Credits

• Base model: DeepSeek-AI
• GGUF quantization foundation: Unsloth
• DeepRiver modification and release engineering: A.I Joe
• Publisher: BlackRiver AI Ltd

See NOTICE.md and LICENSE.