Higher-precision current
Flash Pro
104.2 GBTested footprint
393KValidated context
14.06Decode tok/s
BlackRiver AI / Open-weight model research
DeepRiverV4 Flash.
A quantization-aware behavioral derivative of DeepSeek V4 Flash 0731. Two llama.cpp-ready GGUF variants engineered to preserve capability while substantially reducing refusal behavior.
Release candidate
284.3B parameter class
llama.cpp / GGUF
01 / Two variantsOne model.
One model.
Two currents.
Flash prioritizes the largest practical context and the lowest tested footprint. Flash Pro retains more precision through an IQ3_XXS foundation. Both share the same behavioral modification recipe and direct Q8_0 replacement payloads.
DeepRiverV4 Flash
The broad-context edition, built on Unsloth UD-Q2_K_XL. Designed for 128 GB unified-memory systems where context scale matters as much as decode speed.
FoundationUnsloth UD-Q2_K_XL
Shards3 GGUF files
Tested footprint≈ 96.83 GB
Validated context1,048,576 tokens
Measured decode14.158 tok/s
Measured on Ryzen AI Max+ PRO 395 / Radeon 8060S / 128 GB unified memory using llama.cpp Vulkan. Context is a validated launch configuration, not a stated architectural hard limit.
DeepRiverV4 Flash Pro
The higher-precision edition, built on Unsloth UD-IQ3_XXS. It spends additional memory to retain more of the original quantized foundation while preserving nearly identical throughput.
FoundationUnsloth UD-IQ3_XXS
Shards4 GGUF files
Tested footprint≈ 104.20 GB
Validated context393,216 tokens
Measured decode14.062 tok/s
Measured on the same Vulkan test platform and production router path. The Pro label denotes the higher-precision quantization foundation, not a separate base model.
02 / Behavioral A/BCapability in.
Capability in.
Refusal out.
A corrected four-model A/B compared clean and modified Flash and Pro variants through the production llama.cpp router. The screen disabled reasoning for scoring, required real final content and separated harmless over-refusal from classification-only high-risk boundary probes.
6/6
Both modified variants flipped all six high-risk classifier probes from REFUSE in the clean foundations to COMPLY, without asking the models to generate the embedded harmful content.
Clean: 0 / 6 complyDeepRiver: 6 / 6 comply
Corrected capability screen
10 / 10Flash passed every quality task after modification.
Pro quality screen
9 / 10Strict-format score; 10 / 10 by underlying reasoning.
Performance delta
−1.0%Average decode change across both variants; effectively preserved.
Benchmark interpretation: this demonstrates a substantial refusal-boundary shift in a controlled classification screen. It does not prove universal compliance, universal safety, correctness or immunity to future runtime/template changes. Generated outputs remain model-dependent.
03 / MethodChanged with
Changed with
surgical intent.
The model was not globally requantized. Thirty-three attention output tensors were reconstructed from clean BF16 masters, projected against a validated refusal direction, norm-preserved, encoded once as llama.cpp-compatible Q8_0 and inserted into byte-audited GGUF foundations.
W′ = W − λv(vᵀW)
λ = 3.5 · target layers 10–42 · tensor pattern blk.N.attn_output_b.weight · row norms restored after projection.
Recover clean masters
Exactly 33 BF16 tensors were range-downloaded from the clean Unsloth UD-Q8_K_XL source and individually hashed.
Validate quantization identity
A Python Q8_0 encoder reproduced every existing destination tensor byte-for-byte before any edit was permitted.
Project once
The refusal direction was applied in BF16-derived float space with λ 3.5 and per-row norm preservation.
Encode direct Q8_0 payloads
One-pass payloads were selected over fixed-point correction after controlled comparison.
Patch and verify ranges
Only mapped tensor ranges changed. Metadata, tokenizer, chat template, split structure and every non-target byte were independently checked.
04 / IntegrityTrust the
Trust the
artifact.
DeepRiver carries technical provenance inside its GGUF metadata and in external manifests. The foundations remain credited to Unsloth; the behavioral modification, payload recipe and publisher identity are explicit rather than hidden behind a renamed file.
33
Exact target surface
Thirty-three mapped attention-output tensors across layers 10–42. No auxiliary MTP or unrelated tensors were modified.
0×
No global requantization
The clean quantized foundations were preserved. Only direct Q8_0 replacement ranges were written.
SHA
Hash-backed release
Clean masters, payloads, source reconstruction and final target ranges were hashed and cross-verified.
ProvenanceCredits,
Credits,
inside the model.
Created by A.I Joe and published by BlackRiver AI Ltd, with the original quantization credit preserved.
Base modelDeepSeek-V4-Flash-0731
Quantized foundationsUnsloth UD-Q2_K_XL / UD-IQ3_XXS ↗
Behavioral methodA.I Joe · BlackRiver AI Ltd
Direction SHA-2566e4d8a8f3aa9e21795faab2c5b14d29b019acdf2ddbfbd8238430458a5837fe0
Payload hash-of-hashes1ea0b048fdb93b9cbbf585c861f82a1d84c836452e6795cfb47e985eb496a9dd
Runtime targetllama.cpp-compatible split GGUF
The current is almost public.
Model cards, GGUF shards, checksums, tested llama.cpp presets and the complete technical provenance package will be published through BlackRiver AI.