State of knowledge
The current-state synthesis, organised by subject. For history and process, see the individual claims and
docs/methodology-lessons.md. Provisional = confounded or underpowered; treat as a lead, not a conclusion.
Models
Qwen3.5-122B-A10B (MTP)
The incumbent. Decode 21.90 tok/s at empty context, 18.16 at 32k (clm-0025). Passes its
capability guard and holds up on agentic work: 0.545 mean reward on tau2 airline at
n=22 — not the earlier n=3/n=5 figures of 1.00, which were biased toward easy tasks
(clm-0037, superseding clm-0030). On tasks it completes, it is fast: 19 turns / 2.0 min
median to a correct answer, materially quicker than Nemotron on the same tasks (clm-0039,
provisional — cross-model, unpinned simulator, clm-0043). Thinking is a net negative
for this model specifically: OFF solves 5/5 in ~10 min, ON solves 3/5 and deadlocks
indefinitely on the other two (clm-0033). The KV-dequant patch recovers up to +70.3%
decode at 204,800 context, the depth production actually runs at — stock was running at
59% of achievable speed (clm-0022). Cheapest model measured per correct τ² answer: 6.81
Wh f16, 9.48 Wh q8_0 (clm-0042).
Nemotron-3-Super-120B-A12B
Slowest model measured — 17.51 tok/s decode (clm-0025) — but the most reliable
completer with thinking on: 5/5 tau2 airline tasks at reward 1.00 (clm-0035, the
categorical-completion part; the numeric 0.60→1.00 delta itself is retracted,
clm-0036). On reward alone it is statistically indistinguishable from the 122B at scale
(0.625 ±0.237, n=16, vs 0.545 ±0.208, n=22 — provisional, clm-0039, clm-0043), but
needs 37% more turns and 85% more wall-clock to reach the same correct answers, and costs
87% more electricity per correct answer (12.75 Wh vs 6.81 Wh, clm-0042). Runs at IQ4_XS
against the 122B’s Q4_K_M — an unresolved quant confound on every comparison involving it
(clm-0039).
Qwen3.6-35B-A3B
Fastest decode of the field at empty context (51.01 tok/s, 2.3x the 122B, clm-0023), and
holds speed at depth proportionally as well as the 122B does (17% loss to 32k, both
models, clm-0023). Carries no MTP layers, so that lever is unavailable (clm-0026).
Draft-free ngram speculation is nearly worthless on varied prompts here — 1.11x with 29%
CV (clm-0026), far below a community report’s 216 tok/s solo figure (clm-0010,
unreconciled — different quant finetune and prompt regime, clm-0026). The tool-call
grammar ceiling once attributed to this model class does not reproduce; the real
constraint is context size, not the grammar builder (clm-0024, correcting clm-0023).
gpt-oss-120b
Fastest decode of the four at empty context (55.45 tok/s) but the least bandwidth-efficient
relative to its active-parameter count (34-62% realised across the field; gpt-oss itself
62%, the 35B only 34%, clm-0025). Not currently usable: returns stale answers from a
previous request after the first, at --parallel 1, reproduced 6/6 and not explained by
prompt caching (clm-0025). Requires flash attention to run at all. On tau2 with thinking
off it scores 0.00 (total failure); thinking on brings it to 0.60 (clm-0035, categorical
part only — the numeric deltas are not established, clm-0036).
Screened-but-new candidates
Acquired 2026-08-10, not yet benched — one line each:
- Nemotron-Cascade 2 (30B/3B active) — reasoning-per-active-parameter outlier, gold-medal
competitive-programming claims; text-only, vision audit closed (
candidates/cascade2-30b). - Gemma 4 26B-A4B — claimed ~85 tok/s reflex-tier candidate; vision-capable, one mmproj
download from vision-ready (
candidates/gemma4-26b). - GLM-4.7-Flash — agentic/coding lineage, ~40 tok/s reported on ROCm; text-only
(
candidates/glm-47-flash). - LFM2-24B-A2B — the speed-vs-quality probe for the field, large prefill advantage;
text-only (
candidates/lfm2-24b). - Qwen3.6-27B-MTP — dense architecture, where dense-MTP’s larger 1.4-2x speed-up
(against 1.15-1.25x for MoE) applies; regressed to
acquiredpending re-screen on this box (candidates/qwen36-27b-mtp).
Levers
KV quantisation + the dequant patch
Stock quantised KV costs decode speed at depth (up to 56% at 131k) and the cost grows
with depth; a single cherry-picked patch removes it entirely and slightly reverses it
(+70.3% at 204,800, clm-0022). The workload decides whether q8_0 is worth using, not
the flag (clm-0041, clm-0042): on fixed-token throughput work it is now the cheapest
option once patched (85.3 Wh vs f16’s 89.6 Wh at 200k context, clm-0041); on agentic
work with variable turn counts it costs +39% more turns and +39% more energy per correct
answer than f16, unpatched (clm-0038, clm-0042) — because it does more total work, not
because it draws more power. Whether the patch removes the agentic penalty too is
untested and is the highest-priority open run (see below).
Flash attention
Not a hazard on this hardware’s HIP backend — a claimed dense-model prefill cliff (clm-0017,
community, Vulkan) does not reproduce on ROCm, where a dense 27B reaches 214.66 tok/s
prefill at 32k, 5.2x the broken stock-Vulkan figure at 65k (clm-0021). gpt-oss-120b
requires flash attention to run at all (clm-0025).
Speculation (ngram, MTP)
MTP delivers 1.45x on the 122B (clm-0026 citing clm-0022’s sibling work) but is
throughput-negative at batch per one community source (clm-0010). Draft-free ngram
speculation is close to worthless on varied, non-repetitive prompts (1.11-1.17x with
29-38% CV, clm-0026) but strong on repetitive/structured content — 1.87x on 500 IFEval
prompts community-reported (clm-0016), and a further community report of 3x on
code-editing content that remains unreconciled with our own measurement (clm-0010,
clm-0026). Multi-slot ngram pools inflate throughput when prompts are identical across
slots — a measurement hazard, not a capability gain (clm-0016).
Parallel hazard
--parallel 4 destroys long-context needle retrieval on this build — 0/8 against 8/8 at
--parallel 1 — while throughput reports nothing wrong, so no performance benchmark would
catch it unassisted (clm-0019). This rules out simply copying community multi-slot
configurations that report higher aggregate throughput (clm-0026).
Energy
Wall-metered at the smart plug, whole-box (clm-0040, clm-0041, clm-0042). Idle floor
is 10.1 W (idle-aihydra-2026-08). Under τ² inference the 122B draws ~154 W mean,
Nemotron ~167 W (clm-0040). The correct denominator is per-task energy on a matched
task set, not whole-arm totals — arm-level figures overstated cost by roughly 10x because
most of an arm’s wall time was model loading or non-scoring tasks (clm-0042, superseding
clm-0040). Per correct τ² answer: 122B f16 6.81 Wh (0.21 p), 122B q8_0 9.48 Wh
(0.29 p), Nemotron 12.75 Wh (0.39 p) at 30.3 p/kWh grid import (clm-0042). Self-play
overhead (the model generating both agent and user-simulator turns) inflates these figures
by an estimated 20-25%; correcting for it widens the gaps rather than changing the ranking
(clm-0042). On fixed-token throughput work the picture inverts: the KV-dequant patch cuts
energy 42% at 200k context and makes patched q8_0 the cheapest option (clm-0041) — energy
findings do not transfer between fixed-length and emergent-length workloads.
Open questions
- Does the KV-dequant patch remove the agentic turn penalty, not just the speed
penalty? The obvious next run after
clm-0038/clm-0041/clm-0042— running tonight. - Cross-model comparisons (122B vs Nemotron, thinking on/off) need re-running under a
pinned user simulator. Currently confounded by self-play (
clm-0043); seedocs/methodology-lessons.md§3 for the fix. - A clean Vulkan-vs-HIP comparison on identical commits. A community report shows
Vulkan 55% faster decode at depth, but on a 43-build gap with unsubmitted patches on one
side only — not attributable to the backend as reported (
clm-0031, confidence lowered to low on this basis). - A routing benchmark does not exist. A 12B orchestrator was observed routing
competently in an unrelated project, but nothing in this project’s coverage matrix
measures the routing/orchestrator role (
clm-0034,docs/coverage.md).