Independent benchmarks will determine whether Mach-1 Additive 35B can fit in roughly 7GB and sustain up to 120 tokens per second on consumer or edge hardware while retaining practically useful model quality.

state: expiredheat: lowuncertainty: highknownscott: mediumlocal-inference open-models inference-economicsSyzygyResearch

What is this?

SyzygyResearch has published Mach-1 Additive 35B as a GGUF model on Hugging Face and, in its launch testimony, describes it as using 1.7 bits per weight, fitting in roughly 7GB, reaching up to 120 tokens per second, and retaining 95% of the original model’s quality. The supplied independent snippets show that local-inference performance varies substantially with hardware, context length, and quantization, but none directly benchmarks Mach-1 Additive 35B. Its headline size, speed, and quality-retention claims therefore remain unverified by the provided material.

Why it matters to Scott

Scott already holds the operative position in Capability Audit and hardware-aware local inference: size, throughput, and quality claims must be tested under representative hardware and workloads rather than accepted from launch testimony. Mach-1 could materially expand what his gamepc local-model substrate can run if independently validated, but the supplied evidence currently adds only an unverified candidate, not a new conclusion.

ip:concept.capability-auditdev:concept.hardware-aware-local-inferencedev:project.gamepcip:concept.quality-velocityradar:concept.extreme-quantizationradar:concept.local-inferenceradar:concept.model-evaluationradar:bonsai-extreme-quantization

queries asked of Scott's wikis

ultra-low-bit quantization quality tradeoffs
local inference economics and consumer hardware
benchmarking sustained throughput versus headline tokens per second
open-weight models and edge deployment strategy
memory bandwidth constraints for local LLM inference
quality-retention claims and independent model evaluation

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 (2) — ⭐ canonical anchor

source	object	author	score	comments
🟠 reddit	SyzygyResearch/Mach-1-Additive-35B-GGUF · Hugging Face LocalLLaMA	pmttyji	26	25
🟧 echo.x ⭐	Official launch post: “Today, we're introducing Mach-1 Additive,” described as a 35B model using 1.7 bits per weight, claiming 95% retention	Syzygy Research	—	—

Interpretation history

2026-08-27T20:42:50Z
No independent benchmark or reproducible implementation appeared during repeated observation windows, and the discussion has fully stalled. Retire routine monitoring; a future hardware-specific evaluation would constitute a fresh episode.
2026-08-25T20:34:54Z
The case has gone quiet without producing an independent benchmark, hardware-specific measurement, or implementation report. The artifact remains testable, but repeated staleness now suggests monitoring only for substantive third-party results rather than routine discussion refreshes.
2026-08-23T20:29:42Z
The refreshed comments remain repetitive requests and unsupported usability impressions, not independent measurements. Mach-1 is still a testable artifact, but its headline memory, speed, and quality claims remain wholly dependent on first-party testimony.
2026-08-22T19:39:02Z
The expanded discussion remains repetitive amplification and requests for testing, with only unsupported complaints about speed and CPU usability. No reproducible benchmark or independent implementation changes the core size, throughput, or retained-quality claims.
2026-08-21T19:29:51Z
New comments introduce weak negative impressions about speed and CPU usability, but provide no hardware details, measurements, or reproducible tests. The case remains an independently unvalidated artifact awaiting representative benchmarks of memory use, throughput, and retained quality.
2026-08-20T10:37:46Z
The refreshed discussion still asks for third-party testing and practical tool-calling results rather than supplying them. No independent evidence changes the model’s claimed size, throughput, or quality retention, so the case remains an unvalidated but testable artifact.
2026-08-20T08:36:08Z
No independent benchmark or implementation report has arrived; the lone additional comment is engagement-only and does not strengthen the size, throughput, or quality claims. The case remains a testable artifact awaiting representative hardware results, but there is no reason for near-term attention.
2026-08-20T08:28:12Z
grounded: known/medium — Scott already holds the operative position in Capability Audit and hardware-aware local inference: size, throughput, and quality claims must be tested under rep
2026-08-20T08:25:30Z
origin walked (codex/luna, conf 0.94): anchor reddit.post.1vtcz8t -> echo.x.19f14554bd by Syzygy Research
2026-08-20T08:23:56Z
case created — The released GGUF checkpoints and custom llama.cpp fork are concrete artifacts supporting a distinct, testable local-inference claim.

Decision trace

08-28 06:42expireNo independent benchmark or reproducible implementation appeared during repeated observation windows, and the discussion has fully stalled. Retire routine monitoring; a future hardware-specific evalua
08-28 06:42alert_silentThe only delta is staleness and a minor engagement decline; there is no new evidence about memory use, throughput, CPU viability, or retained quality.
08-28 06:42alert_routeThe only delta is staleness and a minor engagement decline; there is no new evidence about memory use, throughput, CPU viability, or retained quality.
08-26 06:34repriceThe case has gone quiet without producing an independent benchmark, hardware-specific measurement, or implementation report. The artifact remains testable, but repeated staleness now suggests monitori
08-26 06:34alert_silentOnly a minor engagement fluctuation occurred; there is no new evidence about memory use, throughput, CPU viability, or retained quality to surface before the next briefing.
08-26 06:34alert_routeOnly a minor engagement fluctuation occurred; there is no new evidence about memory use, throughput, CPU viability, or retained quality to surface before the next briefing.
08-24 06:29repriceThe refreshed comments remain repetitive requests and unsupported usability impressions, not independent measurements. Mach-1 is still a testable artifact, but its headline memory, speed, and quality
08-24 06:29alert_silentThe new discussion supplies no hardware configuration, benchmark, or reproducible implementation result, so there is no consequential delta to surface before the next briefing.
08-24 06:29alert_routeThe new discussion supplies no hardware configuration, benchmark, or reproducible implementation result, so there is no consequential delta to surface before the next briefing.
08-23 05:39repriceThe expanded discussion remains repetitive amplification and requests for testing, with only unsupported complaints about speed and CPU usability. No reproducible benchmark or independent implementati
08-23 05:39alert_silentComment growth adds no measured hardware result or practical evaluation, so it can wait for an independent benchmark or reproducible implementation report.
08-23 05:39alert_routeComment growth adds no measured hardware result or practical evaluation, so it can wait for an independent benchmark or reproducible implementation report.
08-22 05:29repriceNew comments introduce weak negative impressions about speed and CPU usability, but provide no hardware details, measurements, or reproducible tests. The case remains an independently unvalidated arti
08-22 05:29alert_silentThe refreshed discussion adds unsupported usability complaints rather than a consequential benchmark or implementation result, so it can wait for reproducible third-party testing.
08-22 05:29alert_routeThe refreshed discussion adds unsupported usability complaints rather than a consequential benchmark or implementation result, so it can wait for reproducible third-party testing.
08-20 23:21sensor_dirtyengagement_update
08-20 20:37repriceThe refreshed discussion still asks for third-party testing and practical tool-calling results rather than supplying them. No independent evidence changes the model’s claimed size, throughput, or qual
08-20 20:37alert_silentThe comment refresh adds requests for validation, not validation itself; it can wait until an independent hardware benchmark or real-world implementation appears.
08-20 20:37alert_routeThe comment refresh adds requests for validation, not validation itself; it can wait until an independent hardware benchmark or real-world implementation appears.
08-20 19:21sensor_dirtycomment_update
08-20 18:36repriceNo independent benchmark or implementation report has arrived; the lone additional comment is engagement-only and does not strengthen the size, throughput, or quality claims. The case remains a testab
08-20 18:36alert_silentThere is no consequential new delta beyond a minor engagement change; the release was already routed and its core capability claims remain unvalidated, so this can wait for an independent benchmark.
08-20 18:36alert_routeThere is no consequential new delta beyond a minor engagement change; the release was already routed and its core capability claims remain unvalidated, so this can wait for an independent benchmark.
08-20 18:33alert_shadowThe downloadable model artifacts and inference fork are now available, creating a concrete candidate for Scott’s local-model test stack. Syzygy claims 1.7-bit weights, up to 120 tokens/s, and 95% benc
08-20 18:33alert_routeThe downloadable model artifacts and inference fork are now available, creating a concrete candidate for Scott’s local-model test stack. Syzygy claims 1.7-bit weights, up to 120 tokens/s, and 95% benc
08-20 18:28groundScott already holds the operative position in Capability Audit and hardware-aware local inference: size, throughput, and quality claims must be tested under representative hardware and workloads rathe
08-20 18:25promote_anchororigin walk conf 0.94
08-20 18:23createThe released GGUF checkpoints and custom llama.cpp fork are concrete artifacts supporting a distinct, testable local-inference claim.