@atomic_chat_hq: MTP speedup Qwen by 2.5x in Atomic Chat Dense vs MoE models on 2x RTX 5090 Qwen3.6 27B: 51 → 117 tps +137% Qwen3.6 35B-…
Summary
Atomic Chat's MTP technique speeds up Qwen dense models by 2.5x and MoE models by 25% on 2x RTX 5090 with zero accuracy loss and ~1 GB extra VRAM, using speculative decoding to draft and verify multiple tokens in one pass.
View Cached Full Text
Cached at: 05/21/26, 08:24 AM
MTP speedup Qwen by 2.5x in Atomic Chat
Dense vs MoE models on 2x RTX 5090 Qwen3.6 27B: 51 → 117 tps +137% Qwen3.6 35B-A3B: 218 → 267 tps +25%
MTP drafts several tokens ahead and verifies them in one pass. The speedup depends on memory moved per pass. Dense 27B reads all 27B params per token, MoE 35B-A3B only reads 3B active. Dense had way more to save by batching.
The baseline tps also differ (218 vs 51) for the same reason from the other side. Token generation is memory-bandwidth bound, and MoE moves ~8x less memory per token, so its baseline is already 4x ahead.
~80% draft acceptance. Zero accuracy loss. ~1 GB extra VRAM.
Open-source code and local AI app – in the comments
Similar Articles
RTX 5080 16GB: Qwen3.6 35B MoE at 128k context — 56 tok/s, and why MTP doesn't help
Detailed benchmarks of Qwen3.6 35B MoE on RTX 5080 16GB show that MTP (Multi-Token Prediction) does not improve inference speed at 128k context due to VRAM constraints; the best configuration is Q4_K_XL without MTP, achieving ~56 tok/s generation at 128k context.
@Snixtp: https://x.com/Snixtp/status/2055734339346768225
A user benchmarks the MTP variant of Qwen3.6 27B against the normal version on a single RTX 3090 using llama.cpp, finding MTP offers up to 2.37x faster generation at long contexts (32k-64k) but with slower prefill and no concurrency support yet.
Got MTP + TurboQuant running — Qwen3.6-27B -- 80+ t/s at 262K context on a single RTX 4090
Developer achieved 80+ t/s inference on Qwen3.6-27B with 262K context on a single RTX 4090 by combining MTP (Multi-Token Prediction) with TurboQuant's lossless KV cache compression, sharing their implementation fork and technical details.
Qwen3.5-27B, Qwen3.5-122B, and Qwen3.6-35B on 4x RTX 3090 — MoEs struggle with strict global rules
A user benchmarks three Qwen models (Qwen3.5-27B dense, Qwen3.5-122B-A10B MoE, Qwen3.6-35B-A3B MoE) on 4x RTX 3090 GPUs under real agentic workloads, finding that MoE models consistently underperform the dense 27B at following strict global rules despite speed advantages, with the Qwen3.6-35B leading in generation throughput.
@rohanpaul_ai: Qwen 3.6 27B on a MacBook Pro M5 Max 64GB hitting 34tokens per sec, locally with atomic[.]chat 90% acceptance rate, i.e…
Qwen 3.6 27B achieves 34 tokens/sec on a MacBook Pro M5 Max 64GB locally with 90% draft acceptance, enabled by TurboQuant, GGUF, and llama.cpp, showcasing a major advancement in laptop-based AI inference.