Strata on a 12 GB gaming GPU

Radeon RX 6700 XT · Ryzen 5 3600 · 32 GB DDR4-3200 · Strata v0.1.45 HIP

Status: completed benchmark evidence. Two full quality-suite models; three supplemental speed/context references; four incompatible variants. Results observed October 17–18, 2026 UTC.

Official Strata benchmarks — unmodified upstream tool

Original tools/strata_bench.py speed measurements, three runs per configuration. Tables show medians; ranges and individual requests remain in the canonical data and original tool output. Context allocations are kept separate. A dash means that workload was not run.

Extended tests use --long; the two primary extended tests also use --clients 2. Two-client requests queued on one engine slot. Qwen MTP requests produced zero drafts; Coder produced drafts, without an MTP on/off comparison. Q3 no-MTP compatibility copies add only zero-MTP metadata; tensor bytes are verified unchanged. These are speed measurements, not correctness scores.

Custom benchmarks — practical quality, coding and retrieval

Qwen3.6 Q4_K_XL leads the custom weighted index by 1.36 points. Coder leads HumanEval+ coding and validated context length. Supplemental quant runs have no quality scores.

The index is a custom v2.1 rubric: 90% applied-task categories and 10% HumanEval+. Applied tasks are 24 fixtures repeated three times, not 72 different tasks. This index is not an official Strata quality benchmark.

Context: allocation versus actual recall

Passing three needle positions does not establish broad long-context accuracy. These are highest tested successes, not exact limits.

Dossier workload: three repetitions of the same fixture

Repeated-fixture cache effects differ between models. These are observed workload times, not a controlled cold-start speed comparison. All six dossier outputs were nonempty; this alone does not establish factual correctness.

Memory, hardware and cost

Measured on Proxmox VE 9.2.21 / Debian 13, Linux 7.0.14-24-pve, in an LXC guest with isolated Docker inference and scoring. 32 GB installed RAM; the guest was limited to 28 GiB and inference to 26 GiB. Sampled cgroup memory includes cache and is not a peak measurement. GPU capacity is 12 GB; a reliable inference-only peak VRAM figure is not established.

RAM: 2 × 16 GB Corsair CMK32GX4M2E3200C16 at configured 3200 MT/s. Rated 16-20-20-38 at 1.35 V; active timings/voltage unverified. Storage: Intel 1 TB NVMe; benchmark disk expanded from 300 to 400 GiB to retain models and finish the last reference.

Local hardware: $0 rental and $0 quant-creation spend. Electricity, internet, idle-run cost and hardware amortization were not measured; total project cost is unknown.

Use-case routing among the two fully scored models

Failures and interpretation limits

• Two models received the full quality suite. Q3_K_M, Q4_K_M and Q3_K_XL received only feasibility, speed and recall tests; their quality is unmeasured here.
• The custom v2.1 applied-task rubric and weighted index are not official Strata quality benchmarks. Official refers to the unmodified upstream speed tool; HumanEval+ is separately scored with EvalPlus.
• HumanEval+ here is one greedy sample per Python task, not a comprehensive software-engineering or end-to-end coding-agent evaluation.
• Coder is a pruned Flash-Next variant; Qwen3.6 is a different model. These are deployment choices, not a controlled quant-only quality comparison.
• Q3 non-MTP compatibility copies add only nextn_predict_layers=0. Original model files are preserved and full tensor-data equality is verified.
• Coder retained one practical transport/engine-stall failure among 72 attempts, and three practical output-limit finishes. Its failure was not replaced with a successful retry.
• MTP requested differs from MTP observed. Qwen3.6 MTP files reported zero drafts offered/accepted in the official measurements; Coder showed active drafts. No on/off ablation was run.
• Official short-chat speed is context-allocation dependent. Compare 32K with 32K and 64K with 64K. Small differences and these few samples do not establish statistical significance.
• A context allocation is distinct from actual exercised prompt length and retrieval validation. Highest tested success is not the exact maximum.
• Three needle positions are a narrow recall test, not a guarantee of all long-context tasks.
• RAM readings are sampled Docker cgroup usage, including counted cache; they are not a measured sustained peak or total PC RAM. VRAM peak is not established.
• The two-client official test used one engine slot and queued requests; it is not two-slot engine concurrency.
• Legacy daily_driver_eligible=false reflects unavailable metadata in the old aggregator; it is not a measured suitability verdict.
• the earlier RTX 5070 Ti system used an RTX 5070 Ti 16 GB, more system RAM, and Strata v0.1.43; do not attribute differences solely to GPU vendor or VRAM.
• RAM rated timings are 16-20-20-38 at 1.35V; active timings and XMP voltage were not verified.
• Dossier timing includes repeated-fixture cache effects. No adaptive-policy timing comparison was run.

IQ3_S and IQ4_XS MTP files failed on unsupported draft-layer formats; IQ3_XXS failed on main-model formats. Genesis-Hermes failed on missing metadata. These were compatibility attempts, not completed benchmarks. Q3 compatibility changes are separately recorded and tensor bytes verified unchanged.

Evidence

Canonical data · Verification · Original official outputs · Official-results PR. Operational archives are private and must not be published. Source links:

• https://github.com/Niko1221/Strata/tree/477ac13cbe8e582c20484871ce459d20d82df09d
• https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF
• https://huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF
• https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF
• https://huggingface.co/LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-GGUF
• https://www.corsair.com/us/en/p/memory/cmk32gx4m2e3200c16/vengeance-lpx-32gb-2-x-16gb-ddr4-dram-3200mhz-c16-memory-kit-black-cmk32gx4m2e3200c16