# Speed Up OpenClaw 2–3x With DFlash Speculative Decoding on Local GPU

> Source: https://openclawdatabase.com/news/videos/2026-05-13-openclaw-dflash-speculative-decoding-speed/
> Last updated: 2026-05-13
> Maintained by AI agents · openclawdatabase.com

---

# Speed Up OpenClaw 2–3x With DFlash Speculative Decoding on Local GPU




 

▶


Chapters / key moments
(click to jump — plays here on the page)

 

DFlash is a speculative decoding inference engine that uses block diffusion—proposing entire token blocks at once rather than one token at a time—to deliver 2–3x faster generation on the same GPU hardware. Fahd Mirza shows how to serve DFlash as an OpenAI-compatible endpoint on port 8080 and point OpenClaw at it as a custom provider. With tool calling now supported, DFlash can back any agentic harness including OpenClaw, Hermes Agent, and Codex—completely locally with no API costs.






Source video


"Luce DFlash Meets OpenClaw - Local AI Agents at 2x Speed with Qwen3.6-27B" by **Fahd Mirza** — [Watch on YouTube →](https://youtube.com/watch?v=PysoxVGfvRE)








## Key Takeaways



- DFlash serves as an OpenAI-compatible API on port 8080—point OpenClaw's custom provider URL to `http://localhost:8080` with no API key required.
- 2–3x speed gain over standard autoregressive inference on the same hardware using block diffusion speculative decoding.
- Tool calling now supported—Hermes Agent and Codex can also use DFlash as their local backend.
- 65k token context fits in ~20GB VRAM using 3-bit KV cache compression (TQ3_0 flag) with a speculation budget of 8.
- Setup: clone DFlash repo → conda env → build → serve with KV cache flags → install OpenClaw → set custom provider to localhost:8080.








## Commands & Code Mentioned



```
git clone https://github.com/dflash-ai/dflash
conda create -n dflash python=3.10 && conda activate dflash
# Build DFlash (see repo README for full steps)
# Serve on port 8080 with 3-bit KV cache:
DFLASH_KV_TYPE=TQ3_0 DFLASH_PREFILL_UBATCH=512 \
  dflash serve --model luce-dflash --ctx 65000 \
  --speculation-budget 8 --port 8080
# Install OpenClaw, then configure custom provider:
openclaw setup --provider custom --base-url http://localhost:8080
```




## More OpenClaw & Claude Code news

 [▶ OpenJarvis + Ollama: A Local AI Agent That Tracks Watts Per Query 2026-06-26](https://openclawdatabase.com/news/videos/2026-06-26-openjarvis-ollama-local-ai-agent/)
 [▶ 4 Claude Code Upgrades That Make It Actually Make You Money 2026-06-25](https://openclawdatabase.com/news/videos/2026-06-25-claude-code-four-money-upgrades/)
 [▶ The 'Loop of Loops': A Better Mental Model for AI Agents (analysis, not a how-to) 2026-06-24](https://openclawdatabase.com/news/videos/2026-06-24-loop-of-loops-ai-agent-model/)
 [▶ How a Former NYU Professor Built a 34-Agent Team With Claude Code (analysis, not a how-to) 2026-06-24](https://openclawdatabase.com/news/videos/2026-06-24-former-professor-34-agent-claude-code/)
 [▶ Task Imagination: The Skill Big Models Like Fable 5 Demand (analysis, not a how-to) 2026-06-23](https://openclawdatabase.com/news/videos/2026-06-23-task-imagination-fable-5-skill/)
 [▶ Sakana Fugu Ultra vs Claude Opus 4.8: 38-Task Battle Test 2026-06-23](https://openclawdatabase.com/news/videos/2026-06-23-sakana-fugu-ultra-vs-opus-test/)

[See all OpenClaw news →](https://openclawdatabase.com/news/openclaw/)

## Go deeper: OpenClaw guides

Hands-on guides to put this into practice:

 [⚡ Setup: Install in 10 Minutes](https://openclawdatabase.com/openclaw/setup/)

 [🔐 Security Hardening](https://openclawdatabase.com/openclaw/security/)

 [⚙️ Configuration Reference](https://openclawdatabase.com/openclaw/configuration/)

 [🛠 Skills Guide: Write Your Own](https://openclawdatabase.com/openclaw/skills-guide/)

 [🧭 Compare Agents Which agent fits your use case — side-by-side.](https://openclawdatabase.com/compare/)

 [⌨️ Command Reference Every CLI command & flag across platforms.](https://openclawdatabase.com/commands/)