Can a 16 GB GPU run Mitsuba with the new HiMitsuba LoRA?
The publisher reports that the tested package fits, but 16 GB alone does not prove compatibility. On , the repository was renamed and added the optional HiMitsuba LoRA. The base Mitsuba files did not change. The LoRA works only with the larger PQ2_0 package, not this page's selected PTQ1_0 file.
Publisher files and measurements checked 2026-10-06. File sizes are exact; VRAM and speed are publisher measurements, not Local AI Stats tests.
It is the smaller PTQ1_0 build. The publisher says HiMitsuba was not tested for PTQ1_0 and is for PQ2_0 only.
Required base for HiMitsuba
Mitsuba-ComfyUI-27B-v1.18-PQ2_0.gguf: 7.32 GB / 6.82 GiB, plus the 0.07 GB LoRA.
Download the exact PQ2_0 base and LoRA revision. Do not attach the LoRA to a different Qwen3.8-27B GGUF.
Vision input
The separate projector is 0.63 GB / 0.59 GiB.
Include it when planning image-input storage and memory; model-weight size alone is incomplete.
Measured VRAM
9.5 GB at 131K context and 12.4 GB at 262K context on an RTX 5090 32 GB, parallel 1, q4_0 KV cache, projector on GPU.
These settings fit below 16 GB in the publisher's measurement. They do not establish driver support, available VRAM or speed on every 16 GB GPU.
Compatibility checklist
Use the publisher's required PrismML llama.cpp fork; the checked card says upstream llama.cpp does not support these formats yet.
Turn reasoning off and use the exact PQ2_0 base, LoRA strength 1.0 and stated context/cache settings before comparing results.
Keep enough headroom for the runtime, display use and other processes. File size is not peak VRAM, and measured fit is not a speed guarantee.
NVIDIA or AMD Radeon?
The publisher now provides separate NVIDIA and AMD launchers. The documented NVIDIA path targets Windows 10/11, and the RTX 5090 figures above come from that path. The AMD Radeon launcher is explicitly beta and requires PrismML HIP b10685. The card lists RDNA 4, RDNA 3 and RDNA 2 (gfx1030), but the publisher says it was not tested on real AMD hardware. Treat the AMD path as an installation option to verify, not proof of compatibility or performance; do not transfer the NVIDIA speed result to Radeon.
The publisher measured decode speed falling from 119 to 103.1 tokens/s when applying the LoRA on the stated RTX 5090 setup. This is a measured result for that configuration, not an estimate for another GPU. Read the pinned model card, test conditions, PQ2_0 base file and HiMitsuba LoRA.
These scores were compiled from the linked sources. These published scores were not independently tested here. Different prompts, precision, thinking budgets and tools can change results.
Mitsuba 10-axis suite · v1.18
60.2 / 100
Publisher's custom 10-axis composite for Mitsuba v1.18. It mixes behavior, task, coding, reading, Japanese/prompt and vision tests, so it is kept separate from standard benchmarks.
Benchmark efficiency: 10.037 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Publisher evaluation of the exact PTQ1_0 v1.18 package with thinking disabled in PrismML llama.cpp b10685 on an RTX 5090. The same unpublished questions and settings were used for Mitsuba PQ2_0, Mitsuba PTQ1_0, Ternary Bonsai 2 PQ2_0 and Qwen3.8-27B BF16.
Custom publisher suite, not an independent or standard benchmark. The composite mixes ten differently scoped axes, including refusal behavior, and must not be compared with MMLU, GPQA or other catalog scores. Vision requires the separate 629,246,976-byte Q8_0 projector. PTQ1_0 and PQ2_0 need the PrismML llama.cpp fork; upstream llama.cpp is not supported. The publisher recommends the larger PQ2_0 file because its current vision score is higher. Apache-2.0 license.
Context and version evidence: The repository and file name identify the selected model. Revision hashes are shown above when recorded. Tested context length, full hardware configuration and run date are not consistently recorded in this snapshot. Only settings explicitly stated in each benchmark record are known.
The 5.998 GB figure covers model weights in this file. It excludes the context cache, runtime, operating system and any separate multimodal files. It is not a measured RAM requirement or a token-generation speed.
These complete configurations pass the initial weight-size check for the selected 6.00 GB file. Allow extra memory for context and software. Check the model’s runtime notes and your inference software before buying.
Listing evidence: 2026-09-25. Amazon.com base price visible in an owner-supplied screenshot reviewed 25 Sep 2026; capture time not supplied. Dated reference, not a live quote. Excludes card signup offers, coupons, tax and shipping.
Listing evidence: 2026-10-03. Apple US launch price announced 22 September 2026: USD 1,699. Dated manufacturer reference, not an Amazon offer; tax and shipping may differ.
Listing evidence: 2026-09-24. GMKtec, EVO-X2 / 64GB / 1TB, recorded 22 Sep 2026; exact ASIN capacity identified 24 Sep. This is a dated reference, not a current Amazon price; tax, shipping and bundle contents may differ.
Memory matches are estimates, not tested compatibility or speed. As an Amazon Associate I earn from qualifying purchases.
Confirm model architecture and quantization support in your runner. The selected weights use 5.998 GB; this excludes context cache, runtime buffers and other programs.
These are the first listed tiers that pass our reserve policy, not measured minimum requirements. System RAM and VRAM remain separate. Use the memory tool to change the reserve for your context and runner.