DEEPSEEK V4.1 FLASH Q4 / 27 SEPTEMBER 2026

Start the expert reads
before the GPU is done.

20.30 to 22.74 steady tok/s at 512 prompt / 512 generated, +12.0% over our published 21 September engine on the same three drives, with byte-identical output on every arm.

What was tested

DeepSeek V4.1 Flash Q4, 518,596,067,328 bytes, SHA-256 a5e2e2c3ada4b2e9…. M5 Max, 128 GB, the internal SSD plus two Thunderbolt 5 NVMe enclosures holding byte-identical verified copies, prefill reads split 10:5:5 and decode reads 10:6:6. Control: the binary built from our published tag v41-router-qualified-20260921 with its profile and 4,200 cached experts, the 20.375 tok/s champion of 21 September. Candidate: the same source tree plus three opt-in engine paths and 4,600 cached experts, profile v41-stack-20260927. Same drives, same model copies, same prompt, same session. This page is not a comparison with upstream ds4; the 15 September page has that.

The levers

Each lever was screened alone with interleaved pairs before the stack was qualified as a whole. A is the lever off, B is on; screens ran on the candidate binary with the 21 September profile unless noted.

Measurement

512 prompt tokens, 512 or 200 generated, greedy, context allocation 4,096, a fresh process and empty expert cache per arm, a physical-device sampler and a swap guard. Every arm must reproduce the published reference output hash for its length (f14f5a1dcb74… at 512, 8182ab832dcc… at 200), keep the requested cache, add no swap, and hold at least 1,600 MHz active GPU clock through decode as recorded by an independent powermetrics collector. No build ran and no second engine process existed during any arm. Prompt processing is prefill tokens divided by prefill seconds; generation excludes prefill and startup; steady also excludes the first decode step.

512 generated, BAAB: steady 20.30 → 22.74 tok/s (+12.02%, pairs +2.39 and +2.49), generation 20.04 → 22.33 (+11.40%). 200 generated, ABBA: steady 19.96 → 22.49 (+12.65%, pairs +2.56 and +2.49), generation 19.31 → 21.50 (+11.31%). Against the medians published on 21 September, 20.375 steady and 20.115 generation-inclusive, the new medians are +11.6% and +11.0%. All eight arms produced the reference output for their length.

Every other arm

The single-lever screens and the first attempt at the qualification, which the clock gate stopped: the published binary's control arms ran at 762 and 907 MHz, a low-clock mode this machine enters occasionally and whose cause is unestablished. The passing groups above were rerun after a probe arm held 1,620 MHz; nothing in the source or the settings changed between the attempts, and the faster stack arm from the rejected group was not substituted.

What did not work

Caches of 4,750 and 4,900 experts exceeded the swap-growth guard on this 128 GB machine, with and without deferring the growth until the prefill staging pool is released. Turning the CPU keep-alive off on top of the stack measured −0.09%, inside the noise, so the setting stays. Memory-touching keep-alive kernels, read-pool spinning, completion polling, Q8 row tiling and SIMD-group variants, eighteen reader threads and pipeline pre-warming were each screened with interleaved pairs and rejected; the knobs remain in the source, off by default. Pre-warming experts, staging the release of the prefill pool, static buffer views and warming the slabs did not shorten the first decode step, which remains a separate target.

Against upstream, one to three drives

The stack above was qualified against our own previous engine. To place it against upstream, the same evening ran one interleaved session: pinned upstream ds4 bd66c40 (the base this fork is built on, phase markers only, same toolchain) on the internal SSD, against the fork's v41-stack-20260927 profile on the internal SSD alone, with one enclosure and with two. 512-token prompt, 200 generated, sequence U S1 U S2 U S3 S3 U S2 U S1 U, so every fork rung is bracketed by upstream arms. Upstream sizes its expert cache automatically (3,688); the fork requests 4,600.

The one-drive rung is the portable result: same laptop, same single SSD, same model file, no replicas, 10.18 → 18.66 steady tok/s (1.83×) and 17.96 → 30.98 prompt processing (1.72×). Output is byte-identical to upstream on all 12 arms. Fork arms held 1,620 MHz with zero swap growth; upstream has no keep-alive and its active-weighted GPU clock is recorded rather than gated. The earlier ladder of 15 to 19 September (14.38 on one drive, 18.45 on three, against a 10.6 control) is superseded by this session for the chart on the front page.

28 September follow-up: what four more hours bought

The next morning set out for another 10% and got 1.5%. Two opt-in changes qualified against the 27 September binary built from its tag, same session, same profile and cache, output identical on every arm: a post-MoE flush, which commits a layer's routed work as soon as it is encoded when that layer waited for expert reads so the GPU starts it while the CPU encodes the next layer, and an eviction pre-scan, which ranks cache victims while the CPU is idle on the router mailbox instead of scanning 4,600 entries on the critical path. Profile v41-stack-20260928.

Next-layer prediction, the only lever that could hide the expert-read waits, was re-tested through the flag-readback mailbox and lost in every form: top-2 prediction costs about 1.2 ms of GPU per token and buys 0.85 misses; reading only the predicted non-resident experts cuts misses by a third but the speculative reads share the drives with demand reads and displace resident experts. A token is now about 43 ms: 34 ms of kernels at 69 to 100% of the memory roofline, 5 to 6 ms of exposed read waits, 3 to 5 ms of command-buffer boundaries, 2 ms of CPU. The next real step is a GPU-driven expert gather that removes the per-layer command buffers, which is a multi-day change.

Evidence: the 28 September package, 35 arms with a verifier, release v41-stack-20260928. The drive ladder above was measured with the 27 September profile and is not re-run.

30 September follow-up: prefill

Prefill on this configuration staged only the experts the router selected, which cannot start before the layer’s router has run; the drives therefore idled while the GPU computed the previous layer, about 147 ms of reads alternating with 142 ms of compute per layer. Two opt-in changes, profile v41-stack-20260930: read-ahead staging reads the next layer’s likeliest experts during that window, ranked by a hotlist built from 19 other prompts, and tops up only what did not land; the two-wave routed MoE computes the experts already on hand before the top-up read and the rest after it, bit-exact because every (token, slot) row is written once and the slot sum is a separate fixed-order kernel.

Conditions: the 4,600-expert champion cache tripped the swap-growth guard on a machine with fifteen days of uptime, and the session was run without a reboot; the qualification therefore used 4,200 and the ladder 3,600 on both sides, the ladder accepted swap growth of at most 64 MB per arm (largest observed 0.0 MB, every value recorded), and decode rungs sit below the 27 September ladder, which remains the decode reference. Output matched the published reference hash on all 30 qualification and screen arms and on all 12 ladder arms. Evidence: the 30 September qualification package, the ladder package, release v41-stack-20260930.

4 October follow-up: continuous reader, early Engram reads, conditional prior

Three more read-scheduling changes on top of the 30 September staging, all opt-in and all bit-exact. CPU-side stage marks found that at 512 prompt tokens the two Engram row tables were read synchronously at layers 1 and 14, about half a second each, because upstream overlaps those reads with the sweep only from 1,024 tokens and the drives were busy with expert staging; the profile lowers that threshold and issues the 12,288 random row reads per table from 128 threads so the first table lands inside layer 0. A continuous reader keeps the read-ahead lanes alive across layers instead of starting a thread per layer. And the hotlist is replaced by a conditional prior: how often each expert at layer L+1 followed each expert at layer L in 19 other prompts, applied to this prompt’s own layer-L routing counts (precision of the first 130 experts read ahead on the benchmark prompt: 76.7% → 82.9%; oracle 99.8%).

What did not work: a split router (attention and router in two token halves so the first half’s expert ids start the reads early) was made bit-exact and measured +1.5%: the three-drive prefill is drive-bound (177 GB read per 512-token prompt at the 25.5 GB/s ceiling), so certain reads merely replaced useful guesses while the extra drain lengthened the GPU window. Conditions: the 4,600-expert champion cache tripped the swap-growth guard on a machine with twenty days of uptime, and the session was run without a reboot; every arm here used 3,600 on both sides, the ladder accepted swap growth of at most 64 MB per arm (largest 0.0 MB), and GPU clocks held 1,620 MHz during decode on every fork arm. Qualification package · Ladder package · Release.

Limits

One machine, one model, one prompt at two generated lengths. The stack qualification is against our own published binary in the same session; the ladder is against the pinned upstream base of this fork, not current upstream ds4, and two runs per configuration give an observed range rather than a confidence interval. The 4,600-expert cache needs the headroom of a 128 GB machine. Keep-alive remains enabled and raises power draw. Identical output on every arm is regression evidence, not model-quality validation, and chat or server throughput was not measured.

Download sanitized measurements ↓

Transcribed from the results package in the engine repository, which carries every timing CSV with its hash, the phase accounting and 75 source checksums; a verifier script checks the package without rerunning inference. The JSON excludes machine paths, serials, prompts and model files.

Code. Release v41-stack-20260927 of our fork of antirez/ds4, with the results package and the profile notes. The engine paths are opt-in; the defaults and the earlier profiles are unchanged. Tooling: ArgoDrive.