Independent benchmarks will determine whether dsv4-streaming can run the 284B DeepSeek V4 Flash checkpoint on 64GB Apple Silicon with near-lossless quality and practically useful expert-streaming throughput.

state: expiredheat: lowuncertainty: highknownscott: mediumlocal-inference moe expert-streamingDeepSeekkk-r

What is this?

DeepSeek V4 Flash is described as a 284B-parameter mixture-of-experts model with roughly 13B active parameters, aimed at coding and agentic workloads. The supplied snippets claim it can technically run on 64GB Apple Silicon by streaming experts from SSD, but cache misses sharply reduce generation speed; other material demonstrates or recommends 128GB configurations. The specific dsv4-streaming repository, kk-r’s role, and its claimed cache and near-lossless-quality results are not independently established by these snippets, so the case remains a benchmark claim awaiting verification.

Why it matters to Scott

The radar already tracks the same DeepSeek V4 Flash local-inference validation problem in `radar:deepseek-v4-flash-57gb-local-quant` and `radar:deepseek-v4-nvme-demand-paging`, alongside closely related expert-streaming cases. Verified throughput, cache behavior, and quality measurements would bear directly on Scott’s hardware-aware local-inference policy and evaluation discipline, but the presently unverified repository claim adds no established result yet.

dev:concept.hardware-aware-local-inferenceip:concept.evaluation-driven-developmentradar:deepseek-v4-flash-57gb-local-quantradar:deepseek-v4-nvme-demand-pagingradar:concept.expert-streaming

queries asked of Scott's wikis

MoE expert streaming and cache design
Apple Silicon local inference memory limits
SSD-backed inference throughput economics
quantization quality versus model fit
open-weight frontier models on consumer hardware
local coding-agent model deployment

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (5) — ⭐ canonical anchor

source	object	author	score	comments
🟠 reddit	DeepSeek-V4-Flash (284B) on a 64 GB MacBook: measurements, failures, and a cache result that surprised us LocalLLaMA	cowboy-bebob	2	0
🟧 echo.github ⭐	The linked repository is the primary artifact: it contains the code, raw logs, measurement sweeps, failure analyses, and the cache finding.	kk (GitHub account kk-r)	—	—
🟠 reddit	I ran DeepSeek-V4-Flash (284B) on a 64 GB MacBook. notes and numbers LocalLLaMA	cowboy-bebob	1	3
🟠 reddit	Has anyone run Deepseek V4 Flash on two 128gb Macs using Exo? LocalLLaMA	shveddy	3	9
🟠 reddit	DeepSeek V4 Flash on an M2 Ultra: repacked to 141 GiB losslessly, smaller than the Q4 GGUF, at 25.8 t/s (42 t/s peak) LocalLLaMA	Agusx1211	13	10

Interpretation history

2026-08-25T14:41:13Z
Repeated monitoring has produced only adjacent implementations and anecdotes, with no independent reproduction of the specific 64GB near-lossless throughput claim. The validation episode has faded; a documented matching benchmark should reopen it as a new delta.
2026-08-23T14:31:09Z
The refreshed comments add no measured configuration, reproducible benchmark, or independent validation of the 64GB near-lossless expert-streaming claim. Discussion remains adjacent amplification rather than a new evidentiary line.
2026-08-23T13:36:04Z
A second developer’s Metal optimization branch adds another adjacent implementation path, but no hardware-specific results or reproduction of dsv4-streaming’s 64GB quality and throughput claims. The refreshed discussion therefore broadens the test surface without advancing the central validation case.
2026-08-23T12:34:38Z
The M2 Ultra fork independently strengthens the broader case that hardware-specific repacking and SSD-backed techniques can make DeepSeek V4 Flash practical on Apple Silicon. It uses 192GB unified memory rather than 64GB expert streaming, so it does not corroborate dsv4-streaming’s near-lossless quality or useful-throughput claim.
2026-08-23T12:23:07Z
evidence attached: reddit.post.1vw4qzf — A first-party code artifact demonstrates another practical Apple Silicon path for serving DeepSeek V4 Flash with lossless repacking, SSD-backed KV state, and long context.
2026-08-22T15:28:13Z
The refreshed discussion adds no documented benchmark, configuration, or reproduction beyond the already-known anecdotes about ds4 and multi-machine deployment. The exact 64GB near-lossless expert-streaming claim remains a single-author artifact awaiting independent validation, so the case cools.
2026-08-22T08:29:33Z
A separate user now points to antirez/ds4 as a working, fast single-Mac implementation with newly added two-Mac splitting, strengthening the broader feasibility signal for Apple-Silicon deployment. The report provides no hardware, quantization, throughput, quality, or reliability measurements, so it still does not validate dsv4-streaming’s specific 64GB near-lossless claim.
2026-08-22T00:25:20Z
A separate practitioner now claims 5–8 tok/s on 64GB M1 Max configurations using a similar streaming approach, creating the first plausible independent reproduction lead. The comment lacks logs, methodology, and a clear checkpoint/configuration match, so it does not yet corroborate the repository’s exact quality or throughput claims.
2026-08-21T22:33:00Z
The dual-Mac Exo request broadens the possible deployment test matrix but supplies no implementation or measurements. The core 64GB expert-streaming throughput and quality claims still rest on one author’s artifact and await independent reproduction.
2026-08-21T22:22:33Z
evidence attached: reddit.post.1vutek3 — The request is relevant deployment context for testing whether DeepSeek V4 Flash can be distributed across high-memory Apple Silicon systems, though it provides no benchmark result.
2026-08-21T21:30:59Z
The newly attached Reddit post is same-author redistribution of the same repository and measurements, not an independent benchmark or reproduction. It adds no corroborating line, so the practical throughput and near-lossless-quality claims remain unvalidated.
2026-08-21T21:22:42Z
evidence attached: reddit.post.1vuscye — The linked implementation and reported perplexity and throughput provide valuable independent artifact-level corroboration for expert streaming on 64GB Apple Silicon.
2026-08-21T17:52:08Z
No independent benchmark, reproduction, or implementation has appeared; the only change is weaker engagement with the original unvalidated artifact. The case remains a concrete test target but has cooled pending external verification.
2026-08-21T17:30:54Z
grounded: known/medium — The radar already tracks the same DeepSeek V4 Flash local-inference validation problem in `radar:deepseek-v4-flash-57gb-local-quant` and `radar:deepseek-v4-nvme
2026-08-21T17:28:43Z
origin walked (codex/luna, conf 0.99): anchor reddit.post.1vulskj -> echo.github.0be063915e by kk (GitHub account kk-r)
2026-08-21T17:26:48Z
case created — The released implementation, logs, perplexity measurement, and cache findings define a concrete local-inference claim distinct from existing static-quantization cases.

Decision trace

08-26 00:41expireRepeated monitoring has produced only adjacent implementations and anecdotes, with no independent reproduction of the specific 64GB near-lossless throughput claim. The validation episode has faded; a
08-26 00:41alert_silentThe staleness trigger carries no new consequential fact, and the original claim remains unverified, so there is nothing Scott needs before the next briefing.
08-26 00:41alert_routeThe staleness trigger carries no new consequential fact, and the original claim remains unverified, so there is nothing Scott needs before the next briefing.
08-24 03:21sensor_dirtyengagement_update
08-24 00:31repriceThe refreshed comments add no measured configuration, reproducible benchmark, or independent validation of the 64GB near-lossless expert-streaming claim. Discussion remains adjacent amplification rath
08-24 00:31alert_silentNo consequential new fact has emerged beyond the already-assessed implementation discussion, so this can wait for routine review pending an independent 64GB benchmark.
08-24 00:31alert_routeNo consequential new fact has emerged beyond the already-assessed implementation discussion, so this can wait for routine review pending an independent 64GB benchmark.
08-24 00:21sensor_dirtycomment_update
08-23 23:36repriceA second developer’s Metal optimization branch adds another adjacent implementation path, but no hardware-specific results or reproduction of dsv4-streaming’s 64GB quality and throughput claims. The r
08-23 23:36alert_silentThe new comment links an unbenchmarked optimization branch and requests comparative testing; it establishes neither measured performance nor an independent 64GB reproduction, so it can wait for routin
08-23 23:36alert_routeThe new comment links an unbenchmarked optimization branch and requests comparative testing; it establishes neither measured performance nor an independent 64GB reproduction, so it can wait for routin
08-23 23:21sensor_dirtycomment_update
08-23 22:34repriceThe M2 Ultra fork independently strengthens the broader case that hardware-specific repacking and SSD-backed techniques can make DeepSeek V4 Flash practical on Apple Silicon. It uses 192GB unified mem
08-23 22:34alert_silentThe first-party fork is a substantive adjacent implementation, but its 192GB configuration does not resolve the case’s central 64GB claim and its performance measurements remain unverified; it can wai
08-23 22:34alert_routeThe first-party fork is a substantive adjacent implementation, but its 192GB configuration does not resolve the case’s central 64GB claim and its performance measurements remain unverified; it can wai
08-23 22:23alert_silentA developer has released a hardware-specific llama.cpp fork claiming lossless 141 GiB packaging and 25.8 tok/s on a 192 GB M2 Ultra, but this does not validate the case’s central 64 GB expert-streamin
08-23 22:23surface_candidateA developer has released a hardware-specific llama.cpp fork claiming lossless 141 GiB packaging and 25.8 tok/s on a 192 GB M2 Ultra, but this does not validate the case’s central 64 GB expert-streamin
08-23 22:23alert_routeA developer has released a hardware-specific llama.cpp fork claiming lossless 141 GiB packaging and 25.8 tok/s on a 192 GB M2 Ultra, but this does not validate the case’s central 64 GB expert-streamin
08-23 22:23attachA first-party code artifact demonstrates another practical Apple Silicon path for serving DeepSeek V4 Flash with lossless repacking, SSD-backed KV state, and long context.
08-23 22:22propose_attachA first-party code artifact demonstrates another practical Apple Silicon path for serving DeepSeek V4 Flash with lossless repacking, SSD-backed KV state, and long context.
08-23 01:28repriceThe refreshed discussion adds no documented benchmark, configuration, or reproduction beyond the already-known anecdotes about ds4 and multi-machine deployment. The exact 64GB near-lossless expert-str
08-23 01:28alert_silentNo consequential new capability result occurred; refreshed comments and falling engagement do not justify interrupting Scott before an independently documented benchmark appears.
08-23 01:28alert_routeNo consequential new capability result occurred; refreshed comments and falling engagement do not justify interrupting Scott before an independently documented benchmark appears.
08-23 01:21sensor_dirtycomment_update
08-22 18:29repriceA separate user now points to antirez/ds4 as a working, fast single-Mac implementation with newly added two-Mac splitting, strengthening the broader feasibility signal for Apple-Silicon deployment. Th
08-22 18:29alert_silentThe independent implementation reference is worth tracking but remains anecdotal and does not establish the case’s specific checkpoint, memory, quality, or throughput result; it can wait for repositor
08-22 18:29alert_routeThe independent implementation reference is worth tracking but remains anecdotal and does not establish the case’s specific checkpoint, memory, quality, or throughput result; it can wait for repositor
08-22 18:21sensor_dirtycomment_update
08-22 10:25repriceA separate practitioner now claims 5–8 tok/s on 64GB M1 Max configurations using a similar streaming approach, creating the first plausible independent reproduction lead. The comment lacks logs, metho
08-22 10:25alert_silentThe independent throughput claim is promising but too underspecified to establish a capability change; it can wait for logs, code, or configuration details tying the result to DeepSeek V4 Flash and th
08-22 10:25alert_routeThe independent throughput claim is promising but too underspecified to establish a capability change; it can wait for logs, code, or configuration details tying the result to DeepSeek V4 Flash and th
08-22 10:21sensor_dirtycomment_update
08-22 08:33repriceThe dual-Mac Exo request broadens the possible deployment test matrix but supplies no implementation or measurements. The core 64GB expert-streaming throughput and quality claims still rest on one aut
08-22 08:33alert_silentThe new evidence is only a request for datapoints, with no confirmed run or capability result; it can wait until an independent benchmark or distributed implementation appears.
08-22 08:33alert_routeThe new evidence is only a request for datapoints, with no confirmed run or capability result; it can wait until an independent benchmark or distributed implementation appears.
08-22 08:23alert_silentThe new Reddit post only asks whether anyone has run the model across two 128 GB Macs with Exo; it supplies no measurements, implementation artifact, or confirmed capability change. The existing singl
08-22 08:23alert_routeThe new Reddit post only asks whether anyone has run the model across two 128 GB Macs with Exo; it supplies no measurements, implementation artifact, or confirmed capability change. The existing singl
08-22 08:22attachThe request is relevant deployment context for testing whether DeepSeek V4 Flash can be distributed across high-memory Apple Silicon systems, though it provides no benchmark result.
08-22 08:22propose_attachThe request is relevant deployment context for testing whether DeepSeek V4 Flash can be distributed across high-memory Apple Silicon systems, though it provides no benchmark result.
08-22 07:30repriceThe newly attached Reddit post is same-author redistribution of the same repository and measurements, not an independent benchmark or reproduction. It adds no corroborating line, so the practical thro