Gorgon Halo runs 763.2B DeepSeek at 43.2 Tok/s

Can Gorgon Halo run DeepSeek V4.1 Flash 1017?

Yes, with the right quant and a large enough page file. A Project Maya user report from 29 October 2026 runs DeepSeek V4.1 Flash 1017 on a 192 GB Gorgon Halo at a mean 43.2 tokens a second over three 256-token greedy answers with thinking off, reads prompts at 453 tokens a second at 8K, holds a 256K context, and keeps all 15,744 of the run's expert slots in VRAM with 34,917 MiB to spare.

The build is orcarouter's Abliterated Mix-128G, not a stock checkpoint. The report is one user's machine, not ours. Those limits matter more than the headline, and the number that changes a buying decision is the 34,917 MiB of headroom, not the 43.2.

What the 43.2 tok/s run actually measured

Project Maya issue #663 is a full --bench and --report run, not a screenshot. The machine is one APU with 192 GB of LPDDR5x, a 180 GB page file, and the model files on a Lexar NQ790 4 TB NVMe. The report states that both the GLM and DeepSeek rows fit completely in the GPU's dedicated memory. For DeepSeek, that means all 15,744 experts in VRAM with 34,917 MiB left.

The speed is a mean, not a peak. The bench takes three greedy 256-token answers with thinking off: 46.1 tok/s on a Python prompt, 50.9 on an HTML prompt, 32.6 on prose, and a 43.2 mean. Prompt reading is 453 tok/s at 8K, or 278 at 2K. DSpark is the drafter, and the setup's calibration kept it: 31.6 tok/s with it against 26.9 without.

That is a real result, and it is not a free one. The quant is a community build with refusals removed: 224.57 GB of backbone files plus 161.28 GB of Engram tables, not the full checkpoint's 510.3 GB of FP8 weights. Its resident need is 114 GiB of VRAM and RAM combined. The checkpoint's Engram tables are 196.6B parameters; this build reads Maya's Q6_K copies, 161.28 GB, from the SSD. The 160 GB GPU carve-out holds the backbone; the table does not have to.

Why 192 GB matters more than 273 GB/s

The two machines have the same rated bandwidth, so a bandwidth win is not the claim. The difference is capacity: 64 GB more memory, and a 160 GB GPU carve-out against a usable pool of 112 to 119 GiB.

That gap is the whole story for this model. The Abliterated Mix-128G build is 224.57 GB of backbone files plus 161.28 GB of Engram tables, about 386 GB on disk; its resident need is 114 GiB of VRAM and RAM combined, and the tables are read from the SSD. The report's own memory line is the measurement: 159.83 GB given to the GPU, 6.06 GB of weights, 115.51 GB of experts, 0.24 GB of caches, and 34,917 MiB free. A 128 GB Spark has less room for the same build: the setup's 114 GiB need sits at the top of the Spark's 112 to 119 GiB usable range, while the Gorgon Halo leaves 34,917 MiB.

The point is not that one box is faster. The point is that the 192 GB box can hold the working set resident, and a 128 GB box has to make a different trade.

The 53.2 tok/s screenshot, and what it does not say

A newer post from @PeasantSmith on 31 October at 23:53 UTC shows a higher instantaneous number: the monitor reads 53.2 tok/s while writing, 505 tok/s prefill, 57.6 peak, 134,692 of 1,048,576 tokens, DSpark + Lookup, 100% VRAM hit. The same panel shows 22,716 tokens written over 15 min 39 s, a session average of 24.2 tok/s.

It is a useful signal and a weak record. The post names no quant, no Maya version and no engine build. It may be a newer release, a different quant or a warmer cache. None of that is in the screenshot, so 53.2 is not a second controlled run and it is not the headline. The 43.2 mean in issue #663 is the number with a version, a method and a machine attached.

What the machine actually is

Gorgon Halo is AMD's former codename for the Ryzen AI Max+ PRO 495: 16 Zen 5 cores, 32 threads, a Radeon 8065S with 40 CUs at 3000 MHz, 192 GB of 256-bit LPDDR5x at 8533 MT/s, and 273 GB/s. AMD lists a 55 W default TDP with a 45 to 120 W configurable range, and the Ryzen AI Halo platform at 120 W. The platform page gives Linux or Windows 11; the product page lists Windows 11 64-bit, RHEL and Ubuntu.

Neither AMD product page we read lists the 160 GB carve-out; that figure is the BIOS setting in issue #663, where Windows sees 31.7 GB of RAM and HIP reports 172 GB with the shared part. Our Maya calculator now lists the box with that 160 GB preset, labelled as a reported figure rather than an AMD specification.

The catches

The quant is not stock. Abliterated Mix-128G has its refusals removed. Our 1017 facts sheet has no quality figure for it, so the 43.2 is a speed result for a community build, not a quality claim about DeepSeek's checkpoint.

Windows wants a large page file. With 160 GB given to the GPU, Windows has about 32 GB of RAM left and counts the GPU's share against the page file. The report used 180 GB. On a 128 GB Spark the unified pool is smaller, but the accounting is different.

AMD is experimental in Project Maya. The README lists Gorgon Halo on the AMD path with an experimental label. The report is Windows 11. Linux is listed by AMD, but this run does not test it.

We have not found a matched Spark test. The rated bandwidth is identical. The quant, engine, context and machine differ. Anyone who tells you the 43.2 proves Gorgon Halo is faster than a DGX Spark has not run the comparison.

What it means if you are buying

If the build you want fits in 128 GB, the DGX Spark is the better-supported route, with CUDA and a 200B-parameter guidance line. If your build needs more than that, the 192 GB box is the point. The RTX Spark vs AMD 192 GB page covers that split; the DeepSeek V4.1 Flash 1017 page covers the model and the 763.2B total.

For this specific build, the Maya calculator shows the fit but holds the per-card split, because the abliterated row's tensor sizes have not been read from its headers yet. That is a tool limit, not a memory result. The 34,917 MiB free in issue #663 is the stronger evidence.

The wider lesson is the one the DeepSeek VRAM page keeps making: for a 763.2B mixture-of-experts model, the question is not whether the whole checkpoint fits. It is which experts have to be resident and which table can stay on the SSD. The Strata Qwen3.8 Flash Next page covers the same table-on-SSD trade for a different model.

FAQ

Can Gorgon Halo run DeepSeek V4.1 Flash 1017?

Yes, with a quantized build and enough page file. A Project Maya report runs the Abliterated Mix-128G build on a 192 GB Gorgon Halo at a mean 43.2 tok/s, with all 15,744 of the run's expert slots in VRAM and 34,917 MiB free. The build is not a stock checkpoint.

How fast is DeepSeek V4.1 Flash 1017 on Gorgon Halo?

One Project Maya user report (#663) measured a mean 43.2 tok/s over three 256-token answers with thinking off: 46.1 on a Python prompt, 50.9 on an HTML prompt and 32.6 on prose. Prompt reading was 453 tok/s at 8K, or 278 at 2K. A newer screenshot shows 53.2 tok/s while writing, with a 24.2 tok/s session average, and names no quant or version.

Gorgon Halo vs DGX Spark for local LLMs?

They have the same rated 273 GB/s. Gorgon Halo has 192 GB of memory and a 160 GB GPU carve-out; the DGX Spark has 128 GB and 112 to 119 GiB usable. The Gorgon Halo wins on capacity, not bandwidth, and we have not found a matched test.

Does Gorgon Halo run on Linux?

AMD lists Linux on the platform page, and Project Maya lists Gorgon Halo on its experimental AMD path. The 43.2 tok/s report is Windows 11, so the Linux path is supported on paper but not measured here.

Is 192 GB enough for a 763.2B model?

It is enough for this quantized build, not for the full checkpoint. The checkpoint is 763.2B by the safetensors count, about 748B when the 552B backbone and 196.6B Engram table are added. The Abliterated Mix-128G build needs 114 GiB of VRAM and RAM combined, and Maya's Q6_K Engram tables, 161.28 GB, stay on the SSD.