steadfastgaze/DeepSeek-V4-Flash-0731-2.37bpw-MoEspressoV2

DeepSeek V4 Flash 0731, 2.37 bpw MoEspresso V2 package

This repository holds a MoEspresso package, and MoEspresso is the only engine that serves it: https://github.com/steadfastgaze/MoEspresso.

This package requires MoEspresso 2. Install it with Homebrew:

brew install steadfastgaze/tap/moespresso

llama.cpp, transformers, and mlx-lm will not load these files.

Full serving instructions live in the MoEspresso repository.

Original model: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731

The package includes the DeepSeek DSpark drafter. The quality measurements below use the target model without drafting.

Quality

The first panel compares complete quantized artifacts in their intended runtimes. Lower NLL and perplexity are better; higher first-token agreement is better.

package and runtime	WikiText mean NLL	WikiText perplexity	API-continuation mean NLL	API first-token agreement
MoEspresso 2.37 bpw / MoEspresso	1.71466	5.5548	0.39634	66/100
antirez IQ2_XXS / DS4	1.76824	5.8605	0.41517	54/100
Unsloth UD-IQ2_XXS / llama.cpp	1.71695	5.5675	0.36224	62/100

The WikiText panel uses the first 32 contiguous 2,048-token windows from the held-out WikiText-2 test split and scores 32,736 targets. All three runtimes received identical token IDs. The API panel uses the same 100 prompts and 2,313 provider-selected continuation tokens in every local runtime. Each local model is teacher-forced under that continuation; this is not free-running generation agreement.

MoEspresso was not calibrated on this WikiText test panel. The corpus and separation checks are described below.

The release package also carries these model-specific checks:

check	scope	result
Renderer and tokenizer goldens	34 fixed cases	34/34
Greedy selected-token identity	Five fixed prompts and 14 next-token decisions, compared with the official API	9/14
Official continuation loss	100 prompts and 2,313 target tokens	Average NLL 0.3963; first-token agreement 66/100
Long-context fact recall	16 fixed facts in a 30,000-token-class prompt	16/16 exact match
WikiText all-position release gate	32 windows and 65,504 targets	Package perplexity 6.4774; bf16 teacher perplexity 4.8886

The comparison panel and the all-position release gate score different target positions, so their two WikiText perplexity readings are not directly comparable. Exact artifact identities, corpus hashes, and scoring controls are recorded in the benchmark protocol.

Recipe

The served target model is 2.37 bits per weight. The bundled drafter is excluded from that ratio.

MoEspresso selected IQ2_KS and IQ2_K assignments for the routed experts from calibrated output-error evidence. The dense model weights use q6_K. DeepSeek-V4 auxiliary tensors use the package's declared affine and mxfp8 formats, and the DSpark sidecar uses IQ2_K. The package manifest records every tensor decision and the SHA-256 identity of every file.

After downloading, moespresso-verify <package-directory> checks the complete package against that manifest.

Corpus

The calibration training spine joins two published sources:

• Bartowski's calibration_datav5, assembled from material credited to Dampf and Kalomaze plus edaddario's combined_all_small dataset. It mixes general and instruction-style prose, fenced code, and some multilingual material.
• exllamav3's technical.utf8, which adds technical and scientific prose, including mathematical notation.

The two files are concatenated and tokenized with the DeepSeek-V4-Flash 0731 tokenizer. A deterministic slice is reserved from calibration fitting. Separate C4 prose and source-code files are held out as validation probes. exllamav3's WikiText, multilingual, TinyStories, and synthetic random-token components are not part of the training spine.

The Salesforce/wikitext wikitext-2-raw-v1 test split used for the published perplexity panel was not used to produce importance matrices, calibration moments, or optimizer inputs. A full-file audit found no substantive exact overlap between either calibration component and that test panel or the known historical WikiText train excerpt. This supports the panel-specific statement that the package was not calibrated on the benchmark. It does not establish that every upstream source in the calibration corpus is free of Wikipedia-derived prose, because the complete source genealogy of calibration_datav5 is not known.

Files

The repository contains the target-model safetensors shards, the bundled DSpark sidecar and its shards, tokenizer files, and the package-owned manifests, plans, reports, and runtime sidecars. package_manifest.json is the root of the verification contract.

Credits

The base model and DSpark drafter were trained and released by DeepSeek. The IQ_K formats were designed by Iwan Kawrakow and are served through mlx-iqk. The dense q6_K path uses mlx-kquant. The calibration corpus combines work published by Bartowski and turboderp-org. MoEspresso runs on Apple's MLX.

The complete third-party provenance is recorded in the MoEspresso repository's THIRD-PARTY-NOTICES and in the notices distributed by its dependencies.