rapid-mlx/DeepSeek-V4.1-Flash-DSpark-4d2e-MLX

DeepSeek V4.1 Flash DSpark 4d2e MLX

Experimental three-stage DSpark speculative-decoding sidecar for the Rapid-MLX DeepSeek V4.1 Flash REAP 2-bit target.

This is a sidecar, not a standalone model

This repository contains only the 4.62 GB DSpark head. It cannot generate text by itself and does not contain the approximately 213 GB target model.

Required target:

• Repository: rapid-mlx/DeepSeek-V4.1-Flash-REAP-2bit-MLX
• Qualified target revision: a25fec277b9e7cedc0e9f3f15da874a5cf9d491b
• Runtime: Rapid-MLX's owned DeepSeek V4.1 DSpark implementation
• Recommended verification window: K=4
• Minimum supported machine class: 256 GiB unified memory

Rapid-MLX downloads and verifies this sidecar separately when its experimental DSpark path is enabled. Users who do not enable DSpark do not need these files.

Precision layout

4d2e means:

• Attention, shared-expert, main-projection, Markov, and confidence paths: affine 4-bit, group size 64
• Bandwidth-dominant routed experts: affine 2-bit, group size 64
• Target-shared embedding and LM head: provided by the 2-bit target at runtime

The mixed layout is 4,617,792,648 bytes, only 3.5% larger than the fully 2-bit DSpark head. A uniform 4-bit head was not viable inside the qualified 256 GiB memory boundary.

Measured performance

Environment:

• Apple M3 Ultra, 256 GiB unified memory
• Target: REAP12.5 native affine 2-bit checkpoint, approximately 212.93 GB
• Rapid-MLX packed DSpark verification, K=4
• Four prompt domains: code, arithmetic reasoning, JSON-only structured output, and Chinese
• Up to 128 output tokens, two consecutive repeats per prompt

Path	Weighted decode throughput	Relative to AR
Target autoregressive	9.58 tok/s	1.00x
Mixed DSpark, K=4	19.39 tok/s	2.02x
Mixed DSpark, K=5	18.49 tok/s	1.93x

The K=4 result is approximately 20 tok/s for this measured configuration; it is not a guarantee for every prompt, context length, thermal state, or machine. Per-domain K=4 results ranged from 11.87 to 25.93 tok/s. Mean accepted draft length was 1.85 extra tokens per verification block. Peak MLX memory was 218.232 GB.

K=5 accepted more draft tokens but was slower end to end, so K=4 is the qualified fixed-window setting.

Correctness and experimental status

All four K=4 outputs were token-for-token stable across two repetitions. The verifier is target-authoritative: draft tokens are emitted only after target verification, and corrections come from target logits.

Fixed-batch speculative output is not promised to be bitwise identical to a sequential autoregressive run. Low-margin target decisions can differ with the verification batch shape. The current 2-bit target also showed weak or repetitive instruction-following on several English probes. Treat this sidecar as experimental while broader comparative task-quality qualification is in progress.

Artifact integrity

The manifest records the immutable source revisions and shard hashes. Rapid-MLX pins the repository revision and independently validates every required file's size and SHA-256 before loading it.

File	Bytes	SHA-256
dspark-mixed-stage-0.safetensors	1,556,346,856	381a7fbc8758cd86baab55e8f1ae3020e49e2977f067d4269416889c1aee8118
dspark-mixed-stage-1.safetensors	1,512,099,424	2ca7dc77528ebe3220e6e1633733925ca3420e346b4475163e1639541333876e
dspark-mixed-stage-2.safetensors	1,549,346,368	1ed4663f0487e13372e753b2a5e8b760ae8fd43ba661fae891dfdff62384e64b
config.json	138,476	3518b608fd8b90bc518c7d4533c751367fe3c2416aa3a4f7fde40bc6e498f012
model.safetensors.index.json	265,016	70cbf70324c1b8d7d7c3d5f7f522f38d0be6cba673e90430a0f9a1d8b1ab8cff

Sources and license

The dense DSpark paths were converted from the official DeepSeek V4.1 Flash release at revision dba1be0a40aa45a94ad051997016db3960a90277. Routed expert values come from the pinned affine 2-bit source at revision 802f1a00982705d81b79ad1c83aa0ccc0b863ebc.

This repository is distributed under the MIT license. See LICENSE and NOTICE.