This is a no-requantization K2.1-base + K2-dSpark/MTP variant of DeepSeek-V4-Flash-0731. The 43 target/base blocks come from the calibrated-v1 mixed K2/K3 artifact. Only the routed experts in the three integrated dSpark speculative blocks were replaced with their existing calibrated uniform-K2 tensors. All non-routed MTP envelope tensors remain from the K2.1 source artifact.

Every routed expert is present, and each expert family uses one tier consistently for gate_proj, up_proj, and down_proj. The resulting model has 1,103 K3 expert families in the base blocks and uniformly K2 experts in the three dSpark blocks, for 2.093665 bpw across all 11,776 routed-expert families.

The lower-precision MTP experts produced better speculative acceptance on the matched greedy workload used to qualify this control. With the target/base, non-routed weights, serving topology, prompt shape, output work, and fixed-K5 policy held constant on one RTX PRO 6000 plus four DGX Spark nodes:

Making only the routed MTP experts fuzzier improved strict acceptance by 1.69 percentage points, about 6.1% relative, while both runs returned exactly 8,064 of 8,064 requested output tokens. This is a useful reminder that speculative acceptance measures agreement with the quantized target, not standalone draft reconstruction quality. More MTP bits therefore need not improve acceptance; on a particular target trajectory, the noisier draft can agree more often.

This result is workload- and runtime-specific, not a general claim that lower precision always improves speculative decoding. Base quantization still changes the hidden-state trajectory entering dSpark, and other prompts or sampling policies can produce a different ordering. Full benchmark context and the adaptive-depth measurements are in the DS4RT performance report.

The K2 and K3 source quants share the same 1,426-prompt, 1,081,027-token calibration stream. It emphasizes English and Chinese, with code review and rewriting, worked math, reasoning, structured tool calls, multilingual general knowledge, and a broader language tail. Natural top-6 routes supplied expert activations; experts below 1,024 rows received deterministic candidates from router ranks 7-12, followed only when necessary by an explicitly counted isotropic Hessian residual.

The original mixed allocation scored each expert family by its Hessian-weighted K3-over-K2 error reduction multiplied by the K2 replay's natural squared gate mass. This v2 preserves that allocation in all target/base blocks and substitutes calibrated K2 tensors only for the 3 x 256 routed MTP expert families. No weights were requantized during composition.

Quantization was performed layer by layer with quantized-prefix replay. Hessians were accumulated as FP32 activation Gram matrices, transformed and symmetrized in FP64, and consumed by the EXL3 MCG trellis quantizer with sigma_reg = 0.025. The three dSpark blocks were calibrated with all five proposal rows issued jointly.

The checkpoint follows the source DeepSeek V4 Flash architecture: 43 target blocks, 4,096 hidden size, 256 routed experts with top-6 routing, one shared expert, three integrated dSpark speculative blocks, and a configured maximum position length of 1,048,576 tokens.

An inference runtime must support DeepSeek V4, integrated dSpark speculative decoding, and mixed K2/K3 EXL3 with the MCG codebook. The model is a standard, unsliced Hugging Face sharded checkpoint. Use the tokenizer and prompt formatting described by the original model card.

The quantization and mixing pipeline is built on GPTQModel, ExLlamaV3, and DS4RT. The original model and this quantized checkpoint are distributed under the included MIT license.