Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
codingagentic
User reports MTP improves generation speed by 42% but reduces prompt processing by 18%.
MTP requires at least 2GB free VRAM.
Without MTP the run reaches 1020/42 t/s; with MTP it reaches 830/60 t/s.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is shown only when the source reports it; generation speed cannot tell us time to first token. Check the full setup before comparing.
coding
User reports GLM-5.2 Q1_S outperformed Qwen 3.6 27B Q8 in code generation quality despite the lower quant.
Both models run under the pi harness. GLM-5.2 Q1_S ran at ~6 t/s at low context down to 3 t/s at 100k context, while Qwen 3.6 27B Q8 ran at ~60 t/s.
GLM FP via OpenRouter was also tested, but Q1_S held up well.
User compares PI Agent and Opencode with Qwen3.8 27B on an RTX 3090.
Setup is llama-server with the vision module offloaded to RAM.
User reports PI Agent uses fewer tokens, has no 32k output limit, runs faster, does not freeze, and compresses context later at 90k versus 67k.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is shown only when the source reports it; generation speed cannot tell us time to first token. Check the full setup before comparing.
agenticcoding
User reports Qwen3.8-27B Q6 at 60-63 t/s on dual GPUs (RTX 3090 + RTX 3060).
The run covered 20 hours of agentic coding.
User recommends enabling MTP in llama.cpp for GLM-4.5-Air, a 106B MoE with 12B active params, on 3090s.
User notes MTP also works for full GLM-4.5.
User mentions creative writing and RP finetunes and Intellect 3.x.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is shown only when the source reports it; generation speed cannot tell us time to first token. Check the full setup before comparing.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
User reports Qwen3.6-35B-A3B Q6_K on an RTX 3090 in llama.cpp, with prompt processing improving from 564.5 to 1330.0 tok/s (2.36x) while TG stays at about 97.7 t/s.
Setup offloads 8 MoE expert layers to CPU with larger batch sizes, at 64K context and a Q8 KV cache, on a Threadripper PRO 3955WX with about 100 GB DDR4.
Measurements are two repetitions with about 1.6% drift.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
coding
User reports reliable code generation and codebase ingestion with multi-token prediction enabled.
Setup uses 96 GB of total VRAM.
User reports running Qwen3.6-27B as the reasoning layer in a multi-agent orchestrator for 2 weeks, replacing Claude.
Plan generation is good, but tool-call reliability is poor with a 12% format error rate, long-context drift appears past ~14k tokens, and cascade-failure handling is weak.
The user concludes the model is viable as a reasoning layer but not as an execution layer.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
User recommends getting enough GPUs to avoid VRAM hacks, and runs 2x RTX 3090s.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is shown only when the source reports it; generation speed cannot tell us time to first token. Check the full setup before comparing.
creative-writing
User reports Qwen3.6-27B at ~70 t/s on dual RTX 3090s, generating interactive HTML content inline with chat.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
coding
User reports Qwen3.6-27B-MTP-IQ4_KS.gguf at 72.9 t/s decode and 1261 tok/s prefill on an RTX 3090 24GB at 156k context.
Setup is ik_llama.cpp with a q8_0/q8_0 KV cache, MTP, and vision on CPU.
The user also tested llama.cpp and BeeLlama.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
User benchmarks Qwen3.6-27B FP16 on 4x RTX 3090 with vLLM at TP=4.
Setup is a power limit sweep from 200W to unrestricted, with peak efficiency at 220W.
User reports Qwen 3.6 27B works well as a daily driver.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is shown only when the source reports it; generation speed cannot tell us time to first token. Check the full setup before comparing.
coding
User reports MTP (Multi-Token Prediction) GGUF with the am17an commit for faster inference.
Setup uses a Q4_0 KV cache and a speculative draft of 2.
At 100k context, VRAM usage is about 19 GB, and performance degrades above 90k context.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is shown only when the source reports it; generation speed cannot tell us time to first token. Check the full setup before comparing.
tool-usecoding
User reports about 50 to 66 t/s at roughly 218K context on text, narrative and code workloads.
Tool calls with outputs of about 25K tokens now complete without OOM after a Genesis patch fix (PN12).
Throughput is lower than an earlier configuration, but context length and stability are higher.
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
coding
User reports Qwen2.5-Coder-32B at about 28 t/s generation and around 450 t/s prompt processing on a single RTX 3090.
Setup is llama.cpp build b4400 with a Q4_K_M quant at 32k context and a Q8 KV cache.
User calls it solid for autocomplete, but says it occasionally hallucinates imports in multi-file refactors.
NVIDIA RTX 3090 · Luce DFlash (custom C++/CUDA stack on ggml) · 256,000 ctx
prompt processing (prefill):
913.0 tokens/s
quant:
Q4_K_M (GGUF)
kv:
Q4
flash attention:
on
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
codingmath
User reports speculative decoding with DFlash on a single RTX 3090, targeting Qwen3.6-27B Q4_K_M GGUF at ~16 GB with a z-lab Qwen3.6-27B-DFlash bf16 draft at ~3.46 GB.
Setup is DDTree tree-verify with block size 16, budget 22, and greedy verify. The KV cache is compressed to TQ3_0 at 3.5 bpv, roughly 9.7x versus F16, with a 4096-slot ring buffer enabling 256K context in 24 GB. Decode uses sliding-window flash attention with a 2048-token window, and prefill ubatch auto-bumps from 16 to 192 for prompts over 2048 tokens. The endpoint is OpenAI-compatible and CUDA only, with no Metal, ROCm, or multi-GPU support.
Output is bit-identical to autoregressive in AR mode, and the draft matches the z-lab PyTorch reference at cos sim 0.999812.