Fixed RAM. Confirm the CPU with the seller because the title and CPU table disagree.
Listing evidence: 2026-09-24. Amazon.com price shown in an owner-supplied screenshot reviewed 24 Sep 2026; capture time not provided. Excludes shipping, import charges, tax and optional discounts. This is a dated reference, not a live checkout quote.
Weight size is not tested compatibility or speed. Check runtime support and memory for context before buying. Prices are dated references. As an Amazon Associate I earn from qualifying purchases.
How these options are selected
Up to three platforms, with the lowest dated reference price on each. Unknown prices follow known prices. Each memory pool keeps a reserve: 30% (minimum 2 GB) for system or unified memory; 20% (minimum 1 GB) for dedicated VRAM. System RAM is for CPU inference. This check does not establish GPU allocation, software support or context capacity.
Model file and details
Provider
MemTensor
Total parameters
35,111,542,048
Exact file size
70,223,183,936 bytes
Quantization / precision
BF16
Weight repository
MemTensor/MemT-1.0-Nano
Original model repository
MemTensor/MemT-1.0-Nano
Weight revision
30fe0b22422c4cc4c87475f98db5c8baa5c1a902
Model revision
30fe0b22422c4cc4c87475f98db5c8baa5c1a902
Runtime notes
Advanced NVIDIA workstation model: 70.223 GB BF16 across ALL 26 shards plus memt_block.safetensors. Download the complete pinned repository. Official SGLang requires Linux, CUDA and custom memt_sglang; its two-GPU example is not a tested minimum. No GGUF, Ollama, LM Studio or Apple Metal support was verified. Memory-only computer offers below do not establish MemT runtime compatibility. Extra runtime, context and session memory required. Weights Apache-2.0; repository software PolyForm Noncommercial. GPQA uses a distinct noise protocol; see the test notes.
Recorded release date
2026-10-19
Date evidence
Technical report published 19 October 2026; HF repository creation (15 October) is not used as the announcement date. Reviewed 25 October. Native-memory MoE based on Qwen3.6-35B-A3B; download all 27 indexed weight files, configuration, tokenizer and required code. The file link opens the complete pinned repository.
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.
GPQA Diamond (MemT noise)
78.6 / 100
MemT report native-memory noise condition; fixed irrelevant demonstrations, thinking, eight responses per question. Separate from clean GPQA scores.
Benchmark efficiency: 1.119 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: MemT-1.0 report Table 3 and Appendix D.4, native-memory noise setting. Three fixed BigCodeBench-Hard demonstrations (<2,048 tokens) committed once to empty memory; each question reads a fixed state. Thinking enabled, temperature 1.0, top-p 0.95, top-k 20, min-p 0, presence penalty 1.5, repetition penalty 1.0; output budget 81,920 within 262,144-token context. Eight responses per question with fixed option order; aggregate within each question then across questions. Derived accuracy 78.6% from noise backbone 80.1 and MemT delta -1.5 percentage points. Uncertainty not reported.
Publisher technical report, not an independent or peer-reviewed test. Noise-setting result derived from Table 3: 80.1 minus 1.5 = 78.6; not the clean-backbone score and not comparable to a clean GPQA run. The repository labels the family as 36B total / 3B active; exact published checkpoint tensor count is 35,111,542,048 including native-memory weights. Full BF16 download is 70,223,183,936 bytes across 26 shards plus memt_block.safetensors. Weight storage alone is about 65.40 GiB; runtime, KV cache, dynamic session memory and context need extra memory. Official SGLang requires Linux, NVIDIA CUDA, Python 3.11+, custom memt_sglang and unquantized BF16; the example uses two GPUs, not a proven minimum. The HF card also documents custom Transformers with CUDA 13 / PyTorch 2.11.0. No GGUF, Ollama, LM Studio, Apple Metal or consumer GPU speed/fit result was verified. Ordinary 16–64 GB systems cannot hold the full weights with runtime headroom. High-memory multi-GPU NVIDIA workstation candidate only; capacity alone does not establish compatibility. Nano weights Apache-2.0; repository software PolyForm Noncommercial 1.0.0; third-party terms remain applicable.
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 70.223 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.
Advanced NVIDIA workstation model: 70.223 GB BF16 across ALL 26 shards plus memt_block.safetensors. Download the complete pinned repository. Official SGLang requires Linux, CUDA and custom memt_sglang; its two-GPU example is not a tested minimum. No GGUF, Ollama, LM Studio or Apple Metal support was verified. Memory-only computer offers below do not establish MemT runtime compatibility. Extra runtime, context and session memory required. Weights Apache-2.0; repository software PolyForm Noncommercial. GPQA uses a distinct noise protocol; see the test notes. The selected weights use 70.223 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.