Docs › models/qwen35-122b

Qwen3.5-122B-A10B (MTP)

benched · production incumbent · 122B / 10B active, MoE

Verdict

Qwen3.5-122B-A10B-MTP is the production incumbent, and on this hardware it earns the label. It passes its guard clean at every depth tested, and a single cherry-picked KV-dequantisation patch recovers up to 70.3% of the decode speed a stock build was leaving on the table at 204,800 tokens — production’s own context (clm-0022 measured-here high). Its real τ²-bench airline score is 0.545 at n=22, not the 1.00 that early five-task arms reported, which sampled only the easy end of the domain (clm-0037 measured-here medium). On tasks it does complete, it is fast: a median of 19 turns and 2.0 minutes to a correct answer, well ahead of the only other model measured at comparable scale (clm-0039 measured-here medium). Thinking is a net negative for this model specifically — turned on, it deadlocks two of five adversarial tasks rather than solving them, where turning it off solves all five in about ten minutes (clm-0033 measured-here high). Quantised KV is now close to free: the same dequant patch cuts its energy cost 42% on fixed-token throughput work and turns a 39%-turn agentic penalty into roughly 9% (clm-0041 measured-here medium, clm-0045 measured-here medium). At 6.81 Wh per correct τ² answer it is also the cheapest model measured by a wide margin (clm-0042 measured-here medium). The open confound is the user simulator: every comparison against another model ran this model as both agent and simulator, which is disclosed rather than papered over (clm-0043 measured-here high).

Key numbers

Metric Value Claim
Decode @ 0 tokens 21.90 tok/s clm-0022 measured-here high
Decode @ 32,768 tokens (f16 KV) 18.16 tok/s clm-0022 measured-here high
Decode @ 204,800 tokens (q8_0 KV, patched) 9.69 tok/s clm-0022 measured-here high
Decode @ 204,800 tokens (q8_0 KV, stock) 5.69 tok/s clm-0022 measured-here high
τ²-bench airline reward 0.545 ± 0.208 (n=22) clm-0037 measured-here medium
Turns to a correct answer, median 19 turns / 2.0 min clm-0039 measured-here medium
Energy per correct answer (f16 KV) 6.81 Wh / 0.21 p clm-0042 measured-here medium
Thinking, τ² airline OFF 5/5 @ 1.00 · ON 3/5 (2 deadlock) clm-0033 measured-here high

Decode figures are stock UD-Q4_K_M unless noted “patched” (KV-dequant cherry-pick applied). τ² and energy figures are thinking-off, stock build, f16 KV unless noted.

Guard status

Clean, 4/4. Capability guard at depth 8,000 — coherence, tool-call, isolation, needle retrieval — passes in full on the current benchmark config (run-0005, cfg-0006, aihydra, stock commit 3653e6d). No retraction or open guard failure attaches to this model; the failure mode this class of guard is designed to catch (gpt-oss-120b’s stale-answer bug) does not appear here.

Levers that matter for this model

  • KV-dequantisation patch (ce7689f, cherry-picked onto stock 3653e6d) — the single highest-leverage lever on this model. +70.3% decode at 204,800 (clm-0022); cuts fixed-token energy 42% at 200k context and makes patched q8_0 cheaper than f16 (clm-0041); closes the agentic turn penalty from +39% to +9.3% (clm-0045). Adopting it means carrying a second fork divergence alongside the .ckpt sidecar already in production.
  • Thinking — off. On this model, thinking is a pure liability: no measured accuracy gain and total failure on 40% of an adversarial task set (clm-0033). This is model-specific — the same lever helps two of the three other models measured — so it must not be set as a global default.
  • f16 vs q8_0 KV, turn cost — the metric that actually moves. Reward barely shifts (1.000 vs 0.889 on 9 paired tasks) but q8_0 takes +39% more turns on the stock build, which a coarse binary reward metric hides entirely (clm-0038 measured-here medium). Patched, the same paired design shows +9.3% (clm-0045).
  • MTP (draft-mtp speculation) — production’s shipped config runs --spec-type draft-mtp --spec-draft-n-max 6, delivering roughly 1.45x on this model (clm-0026 measured-here high, citing sibling work to clm-0022). The shipped --spec-draft-n-max 6 sits ~18% below the measured optimum of 2-3 — an untuned setting, not a broken one (clm-0022 note).

Energy

Workload KV Wh pence Claim
τ² per correct answer, matched tasks f16 6.81 0.21 p clm-0042 measured-here medium
τ² per correct answer, matched tasks q8_0, stock 9.48 0.29 p clm-0042 measured-here medium
Fixed-token, 200k context f16 89.55–90.43 ~2.7 p clm-0041 measured-here medium
Fixed-token, 200k context q8_0, patched 85.25 2.58 p clm-0041 measured-here medium
Fixed-token, 200k context q8_0, stock 146.08 4.43 p clm-0041 measured-here medium

Wall-metered at the smart plug, whole-box, at 30.3 p/kWh grid import. Idle floor is 10.1 W; under τ² load this model draws ~154 W mean (clm-0040 measured-here low, content/idle-baselines/idle-aihydra-2026-08). The two workloads invert on KV quant for the same reason: fixed-token work pays for wall time, and the patch cuts wall time; agentic work pays for turns, and unpatched q8_0 takes more of them (clm-0041). Self-play overhead (this model generating both agent and user-simulator turns) inflates the per-task figures by an estimated 20–25%; correcting for it widens the gaps rather than changing the ranking (clm-0042).

Open questions for this model

  • Is the patched +9.3% turn residual real or noise? n=11 pairs, one seed — needs repeats before it is a number rather than a direction (clm-0045).
  • Every comparison against another model is provisional. The user simulator was never pinned — this model played both agent and simulator in each arm — so the 122B-vs-Nemotron and thinking-on/off results need re-running under an independent, pinned simulator before they are load-bearing (clm-0043).
  • MTP draft verification may pass corrupted tokens. A community PSA reports this as an unresolved hazard on gfx1151 HIP, “not even a depth of 1 is truly safe.” Production runs MTP on this model; nothing in this project’s records addresses MTP output-correctness directly — a correctness question, not a performance one, flagged for investigation before any production restore (clm-0044 community medium).
  • Vision is one download away, unverified. The parent is natively multimodal and the staged artifact carries the matching MRoPE sectioning, but an open upstream issue reports CLIP-graph operator gaps with this model+mmproj pairing causing OOM on CUDA; unverified on ROCm/Vulkan (candidates/qwen35-122b, upstream issue #21268).
  • --spec-draft-n-max 6 is untuned. The measured optimum is 2-3; the shipped value costs roughly 18% of the available speculation gain and has not been corrected in production (clm-0022 note).

Provenance

  • Build (stock): ggml-org/llama.cpp @ 3653e6d6d547ec763317d9ecd0ace334a7e21359 (2026-08-07), ROCm 7.1.0, -fa on.
  • Build (patched): same base commit, cherry-picking ce7689f (Nathanw1014’s upstream 2a24abc, “CUDA: dequantize KV on load in the tile FA kernel, use it for quantized decode”).
  • Quant: weights UD-Q4_K_M throughout; KV f16 or q8_0/q8_0 as stated per row.
  • Host: aihydra (current benchmark host, ~128 GiB GTT) for every figure above unless noted. Production’s shipped config (cfg-0002) ran on aibeast (96 GiB GTT, fork carrying the .ckpt sidecar, no KV-dequant patch) — that box is currently down (board fault, inc-0005) and its data has been archived.
  • Dates measured: 2026-08-08 through 2026-08-11.
  • Page reflects the record as of: 2026-08-11.