---
title: "#speculative-decoding articles | Hashnode"
description: "Explore 11 articles tagged #speculative-decoding on Hashnode. Tutorials, guides, and how-tos from the community."
url: "https://hashnode.com/tag/speculative-decoding"
---

Tag

# #speculative-decoding

11 posts · 0 followers

[Graham Andanje](https://hashnode.com/@fata-fanaka) · [Sep 9](https://hashnode.com/posts/speculative-decoding/6aa145115890d09320fa9047)

### [Speculative Decoding](https://fata-fanaka.hashnode.dev/speculative-decoding?utm_source=hashnode&utm_medium=feed)

Different speculative decoding approaches

5 min read · [fata-fanaka.hashnode.dev](https://fata-fanaka.hashnode.dev/speculative-decoding?utm_source=hashnode&utm_medium=feed)

[Manu Shukla](https://hashnode.com/@mr-manushukla) · [Aug 21](https://hashnode.com/posts/lfm25-dspark-ships-32x-on-an-h100-and-118x-on-a-macbook-the-20-august-2026-numbers-mobile-teams-shou/6a889f01b66fa1974421eb1e)

### [LFM2.5-DSpark ships 3.2x on an H100 and 1.18x on a MacBook: the 20 August 2026 numbers mobile teams should read](https://ecorpit.hashnode.dev/lfm25-dspark-ships-32x-on-an-h100-and-118x-on-a-macbook-the-20-august-2026-numbers-mobile-teams-shou?utm_source=hashnode&utm_medium=feed)

LFM2.5-DSpark ships 3.2x on an H100 and 1.18x on a MacBook: the 20 August 2026 numbers mobile teams should read Summary. Liquid AI published LFM2.5-DSpark on 20 August 2026: three speculative-decoding

14 min read · [ecorpit.hashnode.dev](https://ecorpit.hashnode.dev/lfm25-dspark-ships-32x-on-an-h100-and-118x-on-a-macbook-the-20-august-2026-numbers-mobile-teams-shou?utm_source=hashnode&utm_medium=feed)

[Manu Shukla](https://hashnode.com/@mr-manushukla) · [Aug 2](https://hashnode.com/posts/vllm-speculative-decoding-in-2026-p-eagle-vs-dflash-vs-dspark-for-faster-cheaper-llm-serving/6a6fceede45372578ef537b2)

### [vLLM Speculative Decoding in 2026: P-EAGLE vs DFlash vs DSpark for Faster, Cheaper LLM Serving](https://ecorpit.hashnode.dev/vllm-speculative-decoding-in-2026-p-eagle-vs-dflash-vs-dspark-for-faster-cheaper-llm-serving?utm_source=hashnode&utm_medium=feed)

vLLM speculative decoding in 2026: P-EAGLE vs DFlash vs DSpark for faster, cheaper LLM serving Summary. On July 28, 2026, the vLLM project and its Speculators library shipped open-source support for t

15 min read · [ecorpit.hashnode.dev](https://ecorpit.hashnode.dev/vllm-speculative-decoding-in-2026-p-eagle-vs-dflash-vs-dspark-for-faster-cheaper-llm-serving?utm_source=hashnode&utm_medium=feed)

[Shaun Liew](https://hashnode.com/@shaunliew20) · [May 18](https://hashnode.com/posts/qwen3-speculative-decoding-on-the-dgx-spark-two-models-four-methods-one-surprising-lesson/6a0a72353104e2aff0ffc755)

### [Qwen3 Speculative Decoding on the DGX Spark: Two Models, Four Methods, One Surprising Lesson](https://shaunliew.hashnode.dev/qwen3-speculative-decoding-on-the-dgx-spark-two-models-four-methods-one-surprising-lesson?utm_source=hashnode&utm_medium=feed)

A beginner's guide to one of the most practical speed techniques for AI inference, tested on real hardware with real numbers.

15 min read · [shaunliew.hashnode.dev](https://shaunliew.hashnode.dev/qwen3-speculative-decoding-on-the-dgx-spark-two-models-four-methods-one-surprising-lesson?utm_source=hashnode&utm_medium=feed)

[Jangwook Kim](https://hashnode.com/@effloow) · [May 11](https://hashnode.com/posts/parse-parallel-prefix-speculative-decoding-guide-2026/6a01c852daef8f41a8e814d7)

### [PARSE: Faster LLM Inference via Parallel Prefix Speculative Decoding](https://effloow.hashnode.dev/parse-parallel-prefix-speculative-decoding-guide-2026?utm_source=hashnode&utm_medium=feed)

Speculative decoding became the standard inference speedup technique through 2024 and 2025. The idea: a small draft model generates a sequence of candidate tokens, and a larger target model verifies them in parallel — accepting the longest valid pref...

5 min read · [effloow.hashnode.dev](https://effloow.hashnode.dev/parse-parallel-prefix-speculative-decoding-guide-2026?utm_source=hashnode&utm_medium=feed)

[Jangwook Kim](https://hashnode.com/@effloow) · [May 9](https://hashnode.com/posts/gemma-4-mtp-multi-token-prediction-inference-guide-2026/69ff25f3aa311c28e328ce89)

### [Gemma 4 MTP Drafters: How Multi-Token Prediction Delivers 2x+ Faster Local Inference](https://effloow.hashnode.dev/gemma-4-mtp-multi-token-prediction-inference-guide-2026?utm_source=hashnode&utm_medium=feed)

On May 5, 2026, Google released Multi-Token Prediction (MTP) drafters for the Gemma 4 family. The headline claim — up to 3x inference speedup — is technically accurate on specific hardware. The more realistic number for most developer setups is 1.7x ...

6 min read · [effloow.hashnode.dev](https://effloow.hashnode.dev/gemma-4-mtp-multi-token-prediction-inference-guide-2026?utm_source=hashnode&utm_medium=feed)

[Jangwook Kim](https://hashnode.com/@effloow) · [May 9](https://hashnode.com/posts/speculative-decoding-adaptive-gamma-speckv-guide-2026/69feb567e183a3ec2a5e7947)

### [SpecKV: Adaptive Speculative Decoding with Dynamic Gamma](https://effloow.hashnode.dev/speculative-decoding-adaptive-gamma-speckv-guide-2026?utm_source=hashnode&utm_medium=feed)

Every production LLM deployment using speculative decoding is likely running a fixed speculation length of γ=4. That number comes from early benchmarks, it has been copy-pasted across blog posts and framework defaults, and almost nobody questions it....

10 min read · [effloow.hashnode.dev](https://effloow.hashnode.dev/speculative-decoding-adaptive-gamma-speckv-guide-2026?utm_source=hashnode&utm_medium=feed)

[강문규](https://hashnode.com/@devsnack) · [May 6](https://hashnode.com/posts/qwen3-6-on-dgx-spark-vllm-nvfp4-dflash-vs-llama-cpp-2x-faster-at-88-104-tok-s/69fb2aa950ecad45331b9b3a)

### [Qwen3.6 on DGX Spark: vLLM + NVFP4 + DFlash vs llama.cpp — 2x Faster at 88–104 tok/s](https://devsnack.hashnode.dev/qwen3-6-on-dgx-spark-vllm-nvfp4-dflash-vs-llama-cpp-2x-faster-at-88-104-tok-s?utm_source=hashnode&utm_medium=feed)

I benchmarked Qwen3.6-35B-A3B on DGX Spark GB10 using llama.cpp (45–50 tok/s) vs vLLM + NVFP4 + DFlash (88–104 tok/s). Full setup guide, docker-compose, troubleshooting, and real-world results.

10 min read · [devsnack.hashnode.dev](https://devsnack.hashnode.dev/qwen3-6-on-dgx-spark-vllm-nvfp4-dflash-vs-llama-cpp-2x-faster-at-88-104-tok-s?utm_source=hashnode&utm_medium=feed)

[강문규](https://hashnode.com/@devsnack) · [May 6](https://hashnode.com/posts/gemma-4-mtp-drafter-on-dgx-spark-2-89x-speedup-for-dense-31b-no-quality-loss/69fb28d650ecad45331a34d9)

### [Gemma 4 MTP Drafter on DGX Spark: 2.89x Speedup for Dense 31B — No Quality Loss](https://devsnack.hashnode.dev/gemma-4-mtp-drafter-on-dgx-spark-2-89x-speedup-for-dense-31b-no-quality-loss?utm_source=hashnode&utm_medium=feed)

An 870 MB drafter model turned Dense 31B from 6.5 → 18.8 tok/s. No model swap, no training, no quality degradation. If you have a DGX Spark, there's no reason not to use this. Key Results Model Fra

8 min read · [devsnack.hashnode.dev](https://devsnack.hashnode.dev/gemma-4-mtp-drafter-on-dgx-spark-2-89x-speedup-for-dense-31b-no-quality-loss?utm_source=hashnode&utm_medium=feed)

[Adam Yang](https://hashnode.com/@adamygn) · [Nov 24, 2025](https://hashnode.com/posts/turning-latency-into-throughput-speculative-decoding-for-the-decentralized-inference/69244a5d5f55d14f0512a90e)

### [Turning Latency into Throughput: Speculative Decoding for the Decentralized Inference](https://gradient.network/blog/turning-latency-into-throughput-speculative-decoding-for-the-decentralized-inference?utm_source=hashnode&utm_medium=feed)

https://arxiv.org/abs/2511.11733 The Latency Wall In centralized inference, speed is mostly a function of compute. You optimize by saturating HBM bandwidth, fusing kernels, and keeping GPUs close to their roofline. In decentralized inference, where...

6 min read · [gradient.network](https://gradient.network/blog/turning-latency-into-throughput-speculative-decoding-for-the-decentralized-inference?utm_source=hashnode&utm_medium=feed)
