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

ini

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.

powershell

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:

text

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.

BlackRiverAI/DeepRiver-V4-Flash-Pro-GGUF

DeepRiver V4 Flash Pro — GGUF