I wanted to link here @whpthomas recipe as it’s the one that gave me the BEST results overall.
I started very happy with DFlash spec decoding recipes but found them inconsistent across long(ish) sessions with context >128k in OpenCode/Claude Code.
Tried his and not only works flawlessly to me in my scenario, but also scales VERY nicely for parallel workflows. I normally have 2 or 3 coding agents running at the same time and with 2, I feel it basically the same as with a single one. Also, I’m seeing way better Average of Draft Acceptance tokens in real-world (also benchmarks) which translates in net-win output tokens.
Edit: Here are my results :
Benchmarks:
Qwen3.6-35B-A3B-FP8
$ ./bench_multi.sh
Active model: qwen/qwen3.6
Parallel streams: 1╔══════════════════════════════════════════════════════════════╗
║ Parallel Benchmark ║
║ Model: qwen/qwen3.6 ║
║ Mon Apr 20 04:54:02 PM CDT 2026 ║
║ Streams: 1 ║
╚══════════════════════════════════════════════════════════════╝── Run 1/2 ──────────────────────────────────────
[Q&A] 256 tokens in 36.20s = 7.0 tok/s (prompt: 23)[Code] 512 tokens in 7.88s = 64.9 tok/s (prompt: 30)
[JSON] 1024 tokens in 14.85s = 68.9 tok/s (prompt: 48)
[Math] 64 tokens in .92s = 69.5 tok/s (prompt: 29)
[LongCode] 2048 tokens in 30.02s = 68.2 tok/s (prompt: 37)── Run 2/2 ──────────────────────────────────────
[Q&A] 256 tokens in 3.81s = 67.1 tok/s (prompt: 23)[Code] 512 tokens in 7.85s = 65.2 tok/s (prompt: 30)
[JSON] 1024 tokens in 14.42s = 71.0 tok/s (prompt: 48)
[Math] 64 tokens in .94s = 68.0 tok/s (prompt: 29)
[LongCode] 2048 tokens in 30.08s = 68.0 tok/s (prompt: 37)=== Done ===
$ ./bench_multi.sh 4
Active model: qwen/qwen3.6
Parallel streams: 4╔══════════════════════════════════════════════════════════════╗
║ Parallel Benchmark ║
║ Model: qwen/qwen3.6 ║
║ Mon Apr 20 04:56:39 PM CDT 2026 ║
║ Streams: 4 ║
╚══════════════════════════════════════════════════════════════╝── Run 1/2 ──────────────────────────────────────
[Q&A/1-4] 2048 tokens in 18.16s = 112.7 tok/s (prompt: 107 across 4 streams)
[Code/1-4] 2048 tokens in 16.31s = 125.5 tok/s (prompt: 124 across 4 streams)
[JSON/1-4] 2048 tokens in 15.19s = 134.8 tok/s (prompt: 194 across 4 streams)
[Math/1-4] 1703 tokens in 10.90s = 156.2 tok/s (prompt: 116 across 4 streams)
[LongCode/1-4] 2048 tokens in 16.26s = 125.9 tok/s (prompt: 161 across 4 streams)── Run 2/2 ──────────────────────────────────────
[Q&A/1-4] 2048 tokens in 16.49s = 124.1 tok/s (prompt: 107 across 4 streams)
[Code/1-4] 2048 tokens in 15.61s = 131.1 tok/s (prompt: 124 across 4 streams)
[JSON/1-4] 2048 tokens in 14.88s = 137.6 tok/s (prompt: 194 across 4 streams)
[Math/1-4] 1703 tokens in 11.18s = 152.3 tok/s (prompt: 116 across 4 streams)
[LongCode/1-4] 2048 tokens in 16.28s = 125.7 tok/s (prompt: 161 across 4 streams)=== Done ===
After the run, my monitor shows high Draft Acceptance and high real Avg Generation tokens (from vLLM output logs).
My 2 parallel benchmark scales ~2x from a single user… Impressive.