763B DeepSeek vs a 6 GB laptop: Strata's planner reports no usable GPU

Strata, a third-party installer for DeepSeek models (Niko1221/Strata), ships a `--plan` mode that computes a run plan and exits without downloading anything. On this 6 GB laptop all three quantization tiers (1.75 / 2.00 / 2.25 bpw) returned the same verdict: no usable graphics card (the floor is 8 GB; docs were measured on 12 GB) and does not fit, with zero experts placed on the GPU. Downloads would be 239.3 / 256.8 / 273.8 GB. The plan is a soft refusal — it does not block startup, but expects the run to be very slow or to run out of memory. No speed was measured here; the body only cites second-hand references.

Video

Chinese哔哩哔哩 / BilibiliOpen on Bilibili
EnglishYouTubeOpen on YouTube

Results: three tiers, the same verdict

Quantization Download / plan reserve GPU verdict Experts on GPU CPU-side experts
1.75 bpw 239.3 / 245.3 GB no usable graphics card + does not fit 0 114.56 GB
2.00 bpw 256.8 / 262.8 GB no usable graphics card + does not fit 0 132.07 GB
2.25 bpw 273.8 / 279.8 GB no usable graphics card + does not fit 0 149.06 GB

Machine: RTX 3060 Laptop, 6 GB VRAM, about 64 GB RAM (about 47 GB visible in WSL); Strata 0.2.7.1 (strata 0.2.7, commit 46c152a1), main engine 0.1.53, DeepSeek engine 0.1.52; all three tiers at the plan's default 8K context.

One oddity: nvidia-smi showed the card nearly empty (44 MiB in use) and the planner still said there was no GPU. In the 10-29 run, 11 plans and 3 setup.py --check calls left VRAM at 2653 MiB before and after, with the same verdict — the problem is not a desktop eating VRAM, it is that 6 GB does not clear the 8 GB floor.

This is a soft refusal: the plan's own text says the model “can still start; expect it to be slow or to run out of memory”. It is not an observed “it won't run” result.

How it was tested

  • No weights downloaded. Only Strata's built-in --plan, which computes a run plan and exits without modifying files (each run ends with Nothing was changed.).
  • Three quantization tiers at the plan's default 8K context: strata launch --plan --model ist-das-1.75bpw --context 8192 (and 2.00 / 2.25).
  • The 10-29 session ran 11 plans and 3 setup.py --check calls with VRAM at 2653 MiB throughout; the 10-31 screen recording showed 44 MiB before and after.
  • Counterfactual: the same fit logic with the VRAM floor lowered to 6 GB, computed without changing any file.

Where the gap is

Requirement Needs This machine Gap
GPU floor 8 GB (docs measured on 12 GB) 6 GB 2 GB short
CPU-side experts (1.75 bpw) 114.56 GB 47 GB in WSL / 64 GB physical 50–68 GB short
Disk (1.75 bpw, plan reserve) 245.3 GB 305.78 GB free on D: fits, with about 60 GB left

With the floor lowered to 6 GB, the counterfactual places only about 0.8 GB of experts on the card and pushes the remaining 110+ GB to RAM and disk; the tool also warns that the dense weights alone need about 4 GB, so the card may stop on startup. The desktop itself already uses about 2.65 GB of VRAM, leaving about 3.4 GB free — less than the ~4 GB the dense weights need, so an OOM at startup is likely (not verified). In the measured case at 2.25 bpw the CPU-side experts reach 149.06 GB and only about 26 GB of D: remains after the download.

On speed, the plan gives no 6 GB number. Strata's own docs list about 4.7 tok/s on an RTX 5070 12 GB + 64 GB and about 8–10 tok/s on an RTX 3060 12 GB + 46 GB — both second-hand; the 6 GB machine would be slower and probably fail to start, but that is inference, not a measurement.

The boundary is therefore simple: to run this model seriously you want at least a 12 GB card. On 6 GB and 8 GB cards, skip this one. If you want to test anyway, run --plan first before clearing 240 GB of disk.

Reproducing it

# Inspect the plan without downloading or modifying anything
strata launch --plan --model ist-das-1.75bpw --context 8192
strata launch --plan --model ist-das-2.00bpw --context 8192
strata launch --plan --model ist-das-2.25bpw --context 8192

VRAM usage is unchanged before and after --plan, so it is safe on a machine that already has Strata and the matching engine installed. On a brand-new machine Strata may fetch the engine first — do not treat it as an offline check.

Limits and sources

  • Only --plan ran: no weights downloaded and no execution, so speed is unmeasured. The “about 0.8 GB of experts on the card, 110+ GB pushed elsewhere” figure is computed with the same fit logic, not a run.
  • The plan's hard floor is 8 GB (docs were measured on a 12 GB card); older public material often says “12 GB required”. This article follows the plan text.
  • Model size: a separate 16.8 GB figure conflicts with the planner's ~4 GB warning. This article does not use the 16.8 GB figure and keeps only the planner's ~4 GB, which refers to the dense weights, not total model size.
  • Speed figures are second-hand: about 4.7 tok/s on an RTX 5070 12 GB + 64 GB; about 8–10 tok/s on an RTX 3060 12 GB + 46 GB (Strata's own tests, not independently reproduced).
  • The 28–34 tokens/s from the 125B episode belongs to a different model and engine and is not used as a speed comparison.
  • Records: 01-实验/2026-10-28_deepseek-v41-flash-763b-feasibility.md, 01-实验/2026-10-31_deepseek763b-plan.md, 03-文稿/2026-10-31_deepseek763b-script-v002-zh.md; plan log F:\BotWork\logs\dsv41-plan-6g-20261029.log; recording D:\AgentSkills\video-factory\视频项目\deepseek763b\素材\dsv41-plan-6g-screen-20261031.mp4.
  • Strata docs: https://github.com/Niko1221/Strata/blob/main/docs/DEEPSEEK41.md.

Evidence & revisions

run_idSourceClaim
DSP-001Tested herefeasibility.md §10.3All three quantizations yield no_card + does not fit and 0 experts on GPU; downloads 239.3 / 256.8 / 273.8 GB
DSP-002Tested heredsv41-plan-6g-20261029.log; dsv41-plan-6g-screen-20261031.mp4VRAM 2653 MiB before and after the 10-29 plans, 44 MiB in the 10-31 recording; verdict unchanged
DSP-003Estimatedfeasibility.md §10.4Counterfactual at a 6 GB floor: about 0.8 GB of experts on card, 113.8 GB CPU-side, dense weights about 4 GB
DSP-004Tested herefeasibility.md §10.3D: has 305.78 GB free; 1.75 bpw needs a 245.3 GB reserve (about 60 GB left); 2.25 bpw leaves about 26 GB

Revisions

  • 2026-11-09 —— First draft (English) added from the handover package; 2026-11-09 revised after independent review to fix Strata attribution, the loadable field, disk arithmetic and the 4 GB contradiction.