Qwen3.6 MTP Turbo KV Runtime Analysis

Curated Hugging Face analysis draft. This report keeps only completed, comparable 64k-context MTP text runs and explicitly excludes interrupted or non-comparable rows.

Workload Shape and Scope

The headline rows are short-form local inference benchmarks, not full software engineering agent runs. Each completed configuration used 21 prompt cases with 10 repeats, for 210 timed request instances per configuration. Across the 12 completed configurations, the curated data contains 2,520 completed timed requests, 70,440 measured prompt tokens, and 1,005,361 measured generated tokens.

The prompt suite is application-specific to coding assistants and technical agents: Python, C++, TypeScript, Rust, SQL, CUDA debugging, FastAPI/Postgres review, security review, JSON transformation, API design, architecture tradeoffs, math, translation, factual question answering, summarization, and short creative generation. The strongest relevance is for compact technical prompts that produce a few hundred output tokens. It should not be generalized to full repository ingestion, multi-turn tool use, cross-hardware performance, or true 10k-token output sessions without additional testing.

See analysis/workload-scope-and-relevance.md for the full workload scope note.

Charts

Exploratory Long-Output Decode Curves

These charts are a separate measurement mode from the 399-token/request headline benchmark. They track cumulative and per-1k-token interval throughput inside longer generations. They are included for qualitative inspection only and excluded from primary claims because sample sizes are small, prompt tokenization overshot the intended 10k input target, and some streams stopped early.

Code-Level Runtime Analysis

The most likely explanation for the QuinsZouls speedup is not Turbo KV alone. Quins was already much faster in the completed q8_0/q8_0 comparison, before changing the requested KV cache type. Code inspection suggests the difference comes from a more specialized and aggressive MTP path: a dedicated qwen35moe_mtp architecture override, a direct target-context MTP hook, and greedy top-1 MTP draft proposals that fill more of the configured two-token draft budget.

The detailed appendix is included at analysis/mtp-runtime-code-comparison.md and links directly to fixed GitHub source commits: QuinsZouls llama-next, clean MTP llama.cpp, and Atomic TurboQuant. The public-facing conclusion should be careful: this identifies a plausible mechanism, not a fully isolated single-cause proof.

Quins Additions vs Original TurboQuant

The original TheTom TurboQuant fork should be treated as a TurboQuant key-value cache runtime, not as a Qwen MTP runtime. The local TheTom binary exposed Turbo cache formats, but its --spec-type help did not expose mtp, draft-mtp, or nextn. Quins adds the missing Qwen multi-token prediction path: a qwen35moe_mtp/qwen35_mtp architecture override, direct llama_set_mtp registration, hidden-state mirroring through handle_mtp_for_ubatch, greedy top-1 MTP draft proposals, and MTP key-value cache synchronization after rejected or partially accepted drafts.

Quins also adds a grouped query attention-aware cache policy. Grouped query attention means many query heads share fewer key/value heads; when that sharing ratio is high, key-cache compression error can affect several attention heads at once. To reduce that quality risk, Quins can auto-upgrade symmetric Turbo key/value requests such as turbo3/turbo3 into asymmetric q8_0/turbo3: safer eight-bit keys with compressed Turbo values. This is why some requested symmetric Turbo rows are reported with actual q8_0/turbo* key/value types.

Paired Comparisons

Completed Configuration Summary

Host Memory Bandwidth Caveat

This workstation was running system memory at 5,200 MT/s during the benchmark. Windows reports two 16 GiB Kingston DIMMs rated at 5,600 MT/s but configured at 5,200 MT/s. Since all headline runs use MoE CPU offload at --n-cpu-moe 30, faster RAM could plausibly improve absolute throughput by reducing host-memory and CPU-to-GPU staging pressure. This report treats that as a hardware caveat, not a measured causal result, because RAM speed was not swept as an independent variable.

Excluded or Unsupported