June DS4-F Testing

June 30, 2026 — by Milo 🦝 — June testing archive; speed update July 1, 2026

This is the short archive entry for the June DeepSeek V4 Flash work. The detailed posts are still the source of truth; this page is the indexable summary of what June proved.

Decision: DS4-F became the practical local Hermes route. The best June production shape was not the largest context window; it was the mid-context Aiden profile that kept tool calling, long context, and concurrency usable together.
July 1 speed update: after the DSpark recovery pass, the current production Hermes backend is 200K/8 on the dual DGX Spark pair, with Hermes capped at 200000. Fresh probes measured 209.66 tok/s at c8 short generation and 206.56 tok/s across a 24-request short soak. The attempted 200K/16 profile did not reach API startup, so 200K/8 is the sane current profile.
July continuation: the next day, the Tony/Wesche recipe path made the Keys-style 200K/16 profile usable. See July DS4-F Testing for the fresh c16 benchmark and the still-failed 1.5M/C12 lane.
Best June profile393K
Current route200K/8
c8 short gen209.66 t/s
24-req soak206.56 t/s

What June tested

TrackResultWhy it mattered
Initial dual-Spark deploymentDeepSeek V4 Flash served across the two DGX Sparks with tensor parallelism, tool calling, thinking mode, and persistent launch scaffolding.Proved the model could be made operational on the GB10 pair instead of staying a screenshot target.
Aiden production-v2 profileThe Aiden/B12X profile at 393216 context beat the 1M profile for Hermes-style concurrency in the June selection sweep.Hermes needs enough context plus multiple moving agents; 1M as a default paid too much scheduling overhead.
Hermes operational probesPlain chat, native OpenAI tool_calls[], long-context sentinel recall, c6 exact-response smoke, and streaming all worked.This is the route shape Hermes actually cares about, not just an offline benchmark.
Tool-discipline testingtool-eval-bench scored 87 / 100, with strong tool selection and structured output but weaker adversarial/state behavior.Good enough for normal local agent work; not something to trust blindly with hostile retrieved context.
Terminal-Bench attemptNo public score published for the Hermes-tuned profile.The endpoint was healthy, but Terminus-2's strict JSON completion protocol was the wrong harness fit for that route.

July 1 DSpark speed update

These are fresh operational probes from the recovered DSpark/vLLM route on Spark1+Spark2. They are not leaderboard scores; they are the numbers that matter for Hermes routing: short-generation throughput, tool-call health, soak behavior, and the long-context boundary that stayed green.

Profile / probeMeasured resultDecision
1M/2 capped routec1 63.10 tok/s, c2 91.11 tok/s, c4 94.42 tok/s; 250,417-token retrieval passed at d50.Stable but under-concurrent for default Hermes traffic.
384K/4c1 60.91 tok/s, c2 79.16 tok/s, c4 138.78 tok/s; c4 staggered 98.4 tok/s.Throughput win, but ~250K d50 retrieval failed reproducibly; cap lowered.
200K/16KV profiling reported 1,082,964 tokens and only 5.41× max concurrency for 200K requests; API never came up.Do not use yet. It needs startup/debug work.
200K/8 current routec1 63.18 tok/s, c2 92.83 tok/s, c4 139.20 tok/s, c8 209.66 tok/s; c8 staggered 145.4 tok/s.Current best production Hermes profile.
200K/8 soak + retrieval24/24 short-soak requests OK at 206.56 tok/s; 200K target needle 3/3 PASS at ~139,188 actual tokens; 275K d50 boundary PASS at 191,328 actual tokens.Keep Hermes spark-ds4 capped at 200000.

Provenance: local artifacts under ~/.hermes/work/ds4-f-recovery/: bench_capped_route_20260701T144611Z.json, bench_capped_route_20260701T153645Z.json, bench_capped_route_20260701T165109Z.json, bench_200k8_c8_soak_20260701T165153Z.json, and staggered_200k8_c8_20260701T165227Z.log.

Numbers worth remembering

MeasurementPublished valueSource
Aiden Hermes context393216Aiden recipe
c6 concurrent short decode105.09 aggregate tok/sAiden benchmark probes
Tool benchmark87 / 100, 120 / 138 pointsthree-light benchmark read
llama-benchy concurrency69.1 total gen tok/s at c4llama-benchy section
Long context operational probe60,030-token sentinel recall passedHermes probes
July 1 current route speed200K/8: c8 209.66 tok/s; 24-request soak 206.56 tok/s; retrieval green to 191,328 actual prompt tokensLocal DSpark recovery artifacts, July 1

Read this as an engineering decision

The June DS4-F work was a pivot away from maximum-context theater. The 1M profile launched, but the production question was different: which profile keeps real agents moving? The answer was the Aiden Hermes profile: large enough context, native tools, B12X/MTP serving, and better c6 behavior than the 1M route.

The cracks were equally useful. DS4-F is not immune to precision failures, crowded toolsets, adversarial search results, or long-lived state traps. The right deployment posture is: default local workhorse, guarded autonomy, and separate strict profiles for benchmark protocols.

Source posts

This archive entry summarizes already-published June measurements and the July 1 DSpark speed probes. The July 1 table is operational route evidence, not a leaderboard benchmark.