DeepSeek V4 Flash / DSpark Technical Record

Exact family semantics, admitted execution and remaining evidence limits.

Up

This record owns current DeepSeek V4 Flash family facts and the exact DSpark target admitted by YVEX. It does not own macro gate state, operator procedure, or release promotion.

Identity and source

Fact	Value
Family	DeepSeek V4
Canonical target	deepseek4-v4-flash-dspark
Source repository	deepseek-ai/DeepSeek-V4-Flash-DSpark
Pinned revision	62af8fffb2f7030cac4de2f0169f5b8d1101b646
Conversation encoder	encoding/encoding_dsv4.py
Conversation encoder SHA-256	bdbd57c132a1b3725042323d02b98b9d1df28e5f388f134399555d041f5055e0
Snapshot shape	48 Safetensors shards; 72,317 indexed tensors
Target topology	43-layer hybrid SWA/CSA/HCA decoder with mHC and MoE
Draft topology	five-position DSpark block conditioned by three target feature taps
Model maximum context	1,048,576 positions
Hidden width / output vocabulary	4,096 / 129,280
Routed experts / selected per row	256 / 6

The external acquisition record binds the exact revision, sidecars, tokenizer, index, every shard header and payload identity, and the ordered aggregate snapshot identity. YVEX source intake independently produces and consumes its canonical yvex.source_manifest.v3 payload manifest before compilation. The operator record is acquisition evidence, not a substitute input schema; a directory name, mutable branch, or local modification time is never source identity.

The superseded deepseek-ai/DeepSeek-V4-Flash snapshot and deepseek4-v4-flash target are not current aliases. Living product surfaces refuse the old target spelling with one migration hint. Historical records and Git retain its provenance.

Target architecture

Layers 0 and 1 use sliding-window attention. Layers 2 through 42 alternate 21 compressed sparse-attention layers at ratio 4 with 20 heavy-compression layers at ratio 128. SWA uses base RoPE without YaRN; compressed classes use the versioned YaRN extension. HCA retains incomplete groups as raw local history and composes raw and compressed representations without cross-representation deduplication.

mHC represents four 4096-wide streams as one 16,384-wide residual state. Its transitions carry 24 by 16,384 mixing geometry, 20 Sinkhorn iterations, pre/post attention transforms, deferred feed-forward transforms, and final collapse before RMS normalization.

Every target layer has one shared expert and 256 routed experts with top-6 selection. The first three layers use token-ID hash routing. The remaining 40 use a learned BF16 router, sqrt-softplus scores, correction bias, deterministic no-aux top-k, and normalized route weights. Generic MoE, runtime, and backend owners execute the family-selected schedule.

DSpark architecture

DSpark proposes a block of five positions. The target execution plan captures the four mHC streams after target layers 40, 41, and 42, averages each tap, and projects the normalized features into three ordered draft stages. Draft queries use noise token 128799 and mutually visible block positions.

The three draft Transformer/MoE stages are distinct from the public configuration's num_nextn_predict_layers = 1. The bundled inference configuration declares n_mtp_layers = 3, and the exact tensor geometry establishes those three executable stages. Stage zero owns the main feature projection and normalization. The final stage owns output normalization, the rank-256 Markov projections, confidence projection, and the final hidden combination. Target embedding and vocabulary output resources are shared by identity rather than copied into a second model.

The Markov component is a low-rank token-conditioned vocabulary bias, not a second persistent recurrent state. Confidence values are scheduling facts; they never replace full-target verification or authorize publication.

Tokenizer and output

The target has a 128,000-entry base tokenizer plus 1,283 added-token records and an untied 129,280-entry output head. The tokenizer owner admits exact prompt rendering, UTF-8 encode/decode, special/EOS classification, incremental detokenization, and committed-token append semantics.

The source declares a one-million-token context contract. Hosted context capacity is selected by the startup profile and admitted by the engine specialization; source capacity is not an automatic runtime configuration or performance claim.

These facts are sealed once by the family projection into model-execution descriptor schema v1. The descriptor also carries RoPE/YaRN, attention-class, compression, indexer, mHC, FFN, output, proposal, feature-source, Markov, confidence, special-token and persistent-state geometry. Common runtime, server, protocol, CLI and generic backend code consume that descriptor rather than duplicate these values.

The pinned conversation encoder owns three admitted modes. Chat/non-think emits the assistant marker followed by </think>; think-high emits the marker followed by <think>; think-max additionally prepends the encoder's exact maximum-effort instruction. These source facts live in the model-family conversation descriptor. Common runtime, protocol and adapters consume typed policies and channels and contain no DeepSeek token literals.

In thinking mode, output before the exact source-authored </think> token is explicit reasoning and output after it is final content. The tokenizer owner consumes the delimiter and refuses an unfinished reasoning grammar. It does not search disabled-mode prose or classify delimiter-looking text in ordinary final content. DSML tool blocks are parsed only through the source grammar and remain a third typed result. This is model-emitted text, not access to hidden internal chain of thought.

Ordinary multi-turn prompts apply the encoder's drop_thinking behavior: reasoning from assistant turns before the latest user turn is omitted. A tool-enabled prompt disables that drop so reasoning, calls and ordered tool results retain the continuity required by the official format.

The independent reference.deepseek.official-encoding gate authenticates the four upstream encoding/parsing cases and gold strings at the exact DSpark revision above, using tests/vectors/manifest.json. It executes the unmodified upstream test, compares native artifact-bound BPE encoding/decoding for all four strings against tokenizers==0.20.3, and compares the supported request prefix of case 2 against the upstream encoder. Thirteen additional text controls and three simple prompt controls remain separate. This is not native projection of every tool/developer/reminder transcript and is not an official full-model logits oracle. Source code, gold files, tokenizer data and their upstream MIT license stay in the immutable external source snapshot; absent or changed backing fails the gate rather than downloading an unpinned substitute. Run make test-deepseek-official-vectors with DEEPSEEK_SOURCE, DEEPSEEK_SELECTED_ARTIFACT, YVEX_RUNTIME_BINDING and the pinned test-only YVEX_TOKENIZER_REFERENCE_PYTHON. No GPU or generation runs.

Coverage, transformation, and artifact

Exact source coverage reconciles 72,317 source tensors, 3,130 more than the superseded snapshot inventory. The sealed Transformation IR has 1,409 terminal descriptors: 1,328 target-trunk descriptors and 81 DSpark descriptors. Every source tensor is required and mapped, explicitly shared, or rejected; scale companions, expert coordinates, target taps, draft stages, and global shared resources remain distinct.

The current bootstrap physical policy is deepseek-v4-flash-dspark-bootstrap-q2-v1. It preserves the admitted mixed IQ2_XXS/Q2_K decisions for the same target roles, not on an assumption that the replacement checkpoint has identical payload bytes. The retained DS4 importance matrix remains bound to its predecessor source identity and is used only as a bootstrap prior; it is not represented as DSpark calibration. New draft norms, controls, feature and Markov projections, confidence tensors, and other sensitive small roles use conservative exact, BF16, or Q8_0 storage. Draft expert decisions are role-specific and cannot inherit an aggressive low-precision default from their auxiliary scope.

The complete GGUF records source, logical model, transformation, physical variant, target and draft role inventories, DSpark configuration, and exact artifact identity. One runtime binding requires target execution, draft execution, complete target verification, persistent target state, and bounded draft workspace. Container validity or target-only opening alone does not establish DSpark support.

Hosted execution

One immutable engine generation owns both target-only and dspark execution plans under its authenticated package and deployment specialization. One engine-bound server session owns committed target state, token ledger, transcript, incremental decoder and sampling state, plus bounded draft and verification candidate state. No second process, model opening, tokenizer, session registry, CUDA context, or output head is created for drafting.

In DSpark mode, proposals do not advance position, KV, transcript, usage, or text. The complete target verifies the ordered candidate block and retains an exact checkpoint after each target-authored row. Greedy mode requires exact target-token equality; admitted stochastic mode uses target-distribution-preserving accept/reject and residual sampling. The runtime promotes only the accepted target-authored checkpoint without replaying accepted rows, discards the rejected suffix, and publishes text after model, token, decoder, and RNG state agree.

Target-only remains the explicit semantic reference and debug mode. An explicit DSpark request fails closed when any draft tensor, plan, qtype, workspace, backend capability, or policy requirement is absent. It never silently falls back to target-only.

For explicit reasoning, the source-authored terminator ends the speculative shape; the final channel continues with ordinary target decode. The source-boundary event reports boundary extent in a, target-only continuation in b, and replayed accepted target rows in c (required zero). This identity-bound sub-policy is not silent fallback.

Current capability

DeepSeek-V4-Flash-DSpark has a complete YVEX source-to-streamed-text path and is the only currently admitted target-verified speculative family. The hosted native, interactive, and bounded OpenAI-compatible paths consume one target-verified runtime authority. Target-only and DSpark modes, multi-turn reuse, cancellation, reset, and committed-only streaming are implemented under the current local protocol. The admitted tokenizer/prompt profile also supports explicit reasoning high, maximum, and disabled policies through separate reasoning, final, tool, and error channels.

Explicit non-claims

This record does not claim:

• an optimized GB10 physical variant or a DSpark speedup;
• native MXFP4/NVFP4 Tensor Core execution;
• production load-aware confidence scheduling or continuous batching;
• complete accelerator residency or an entirely device-side generation path; admitted CUDA greedy/stochastic selection and speculative correction retain common transactional RNG and host tokenizer ownership;
• multi-device or distributed serving;
• model behavior or quality evaluation or quality parity;
• a public full-model benchmark;
• speculative support for another family;
• release qualification.

Current gate state is owned only by Status.