A Drafter Catalog

Targets at a glance

Both models follow the 2026 pattern the Atlas predicted: the drafter ships with the weights, trained alongside the target. GLM-5.2 carries a DeepSeek-style nextn MTP layer, upgraded in this release (IndexShare/KVShare inside the MTP layer, rejection-sampling-aware training). V4-Pro goes further and ships two speculators: the classic 1-layer MTP and DSpark — a semi-autoregressive 5-token block drafter reusing target weights with a low-rank Markov head and confidence-scheduled variable-length verification (arXiv 2607.05147, official code in DeepSpec).

Naming checks: GLM-5.2 has no Air/Flash variant (the small siblings are GLM-5.3-Flash and GLM-4.7-Flash). "DeepSeek-V4-Pro-Max" is a reasoning mode, not separate weights; there is no V4-Lite, and R2 never shipped. Tech reports: GLM-5 (arXiv 2602.15763; GLM-5.2 has no separate report — the 4.56→5.47 MTP figure is Z.ai's release ablation) · DeepSeek-V4 (arXiv 2606.19348) · DSpark (arXiv 2607.05147; paper says Markov rank 256 "by default", shipped V4-Pro configs use 512).

GLM-5.2: use the shipped MTP; DSpark and DFlash are the trained alternatives

First choice — the built-in MTP layer. Zero extra artifacts, production-proven in both engines (Z.ai ablation mean 5.47; LMSYS reports "frequently 5+" and >500 tok/s bs=1 on 8×B300; vLLM's 24×B300 P/D deployment measured ~60% acceptance, mean accept ~3→~4 and TPOT 40→22→17 ms). Code/math-heavy traffic is the exception: UCloud's DFlash block drafter accepts 6.4–8.0 tokens there and beats MTP by 1.42–1.54× end-to-end in SGLang, but falls to 0.93× of MTP on chat and Chinese — so it is a workload-specific upgrade, not a default.

# vLLM ≥ 0.23 — method "mtp"
vllm serve zai-org/GLM-5.2-FP8 --kv-cache-dtype fp8 -tp 8 \
  --speculative-config.method mtp --speculative-config.num_speculative_tokens 5

# SGLang — nextn weights ride the EAGLE algorithm path
python3 -m sglang.launch_server --model-path nvidia/GLM-5.2-NVFP4 -tp 4 \
  --speculative-algorithm EAGLE --speculative-num-steps 5 \
  --speculative-eagle-topk 1 --speculative-num-draft-tokens 6 --kv-cache-dtype fp8_e4m3

DeepSeek-V4-Pro: DSpark first, MTP as the portable fallback

First choice — the shipped DSpark module. The only Pro-sized option with published wins: accepted length ~5 vs ~2.5 for MTP-1, +57–78% per-user speed over MTP-1 on Pro (+60–85% on Flash), +52% aggregate throughput at a 35 tok/s/user SLA and a nominal +406% at 50 tok/s/user where MTP-1 collapses (paper §5.4), 383.7 tok/s bs=1 TP=8 on B300 (SGLang v0.5.16 release notes), and 57.2% acceptance / τ 3.86 on a 2×H100 NVL field report (≈1.75× the per-step yield of MTP at 2.2). No third party has trained a head for Pro — the only community artifacts are GGUF extractions of this same module.

# SGLang ≥ 0.5.16 — DSPARK shipped (#30261, #31434); no separate --speculative-draft-model-path
python3 -m sglang.launch_server --model-path deepseek-ai/DeepSeek-V4-Pro-0813 -tp 8 \
  --speculative-algorithm DSPARK        # env SGLANG_RAGGED_VERIFY_MODE=compact|static|cap-accept
                                        # MTP fallback: --speculative-algorithm EAGLE \
                                        #   --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4

# vLLM ≥ 0.25 (PR #46995) — note V4 uses "mtp", not V3's "deepseek_mtp"; fp8 KV + FLASH_ATTN/sparse-MLA backend required
vllm serve deepseek-ai/DeepSeek-V4-Pro-0813 -tp 8 --tokenizer-mode deepseek_v4 --kv-cache-dtype fp8 \
  --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"probabilistic"}'
#   main (Aug 2026, PR #47808): add "enable_adaptive_verification":true — SM100-class backend only;
#   no PP / LoRA / --enforce-eager / output logprobs in that mode. 0.27.0 adds DSpark AR fusion (#50242).

# llama.cpp / ik_llama.cpp — extracted Pro DSpark drafter GGUF (77B sparse, 26–42 GB quantized)
llama-server -m DeepSeek-V4-Pro-0813-.gguf --spec-type draft-dspark \
  --spec-draft-n-max 5 -md dspark-DeepSeek-V4-Pro-0813.gguf   # pokey2025: ~46% acceptance, 1.3–1.9× on 1×RTX Pro 6000 + CPU-MoE
# (the widely quoted unsloth 1.91× / 76.4% on B200 is the V4-Flash-0731 drafter, not Pro)

Engine support matrix

Decision guide

Training your own head

Three toolchains, all open: DeepSpec (DeepSeek official, MIT) trains and evaluates DSpark / DFlash / EAGLE-3 end-to-end — the reference for V4 targets. Speculators (Red Hat / vLLM; DSpark since v0.7.0) is proven on both of this page's flagships — it produced the GLM-5.2 DSpark heads (RedHat's and siro1's) and the V4-Flash DFlash head, with online hidden-state extraction from a live vLLM server as the documented recipe. SpecForge (LMSYS/SGLang) trained the AQ-MedAI EAGLE-3 heads and, via its offline hidden-state pipeline, UCloud's GLM-5.2 DFlash drafter — so a GLM-5.2 target path exists in practice even though the repo README lists no GLM-5.2 config; DeepSeek-arch target support is tracked in issue #257. z-lab/dflash (the DFlash authors) still lists GLM-5.1 as "coming soon". Rule of thumb from every card surveyed: match the training data to your serving distribution (RedHat's task-spread — 5.27 on math vs 3.14 on QA from one head — and UCloud's 8.04 on MBPP vs 2.98 on C-Eval show how much it matters), and re-benchmark against the shipped MTP/DSpark before adopting any third-party head. Two of the three GLM-5.2 DFlash/DSpark cards state no hardware for their numbers.