Can an 8 GB, 12 GB or 16 GB GPU run Ternary Bonsai 2?
The publisher of a separate MTP graft now reports measured runs on RTX 2070 Super 8 GB, RTX 3060 12 GB and RTX 5060 Ti 16 GB cards. These results are useful compatibility evidence, but they do not apply automatically to this page's selected Ternary-Bonsai-2-27B-PTQ1_0.gguf file. The measured package contains an added multi-token-prediction head and requires a specialized runtime.
The repository had 41,256 downloads and 67 likes when checked on . Those counts show repository activity, not unique users, search volume or independent validation.
Publisher measurements for the exact MTP files and stated runtime; not Local AI Stats tests.
Publisher live-stream greedy decode, 3 x 512 generated tokens, q4_0 K/V, ubatch 256; no display for the 48K result. With a desktop active, the card held 16,384 tokens; 65,536 did not load.
Publisher client probe on bonsai2 v1.1 with batch-invariant mode; 16 GB serve line and vision projector loaded. The publisher recorded 15,070 MiB after load and 15,326 MiB after an image request.
Before downloading or buying hardware
The catalog's selected base file is 5.95 GB / 5.54 GiB. The tested MTP files are 6.30 GB and 7.01 GB because they add the prediction head; do not use the base-file size for the measured setup.
Use PrismML llama.cpp or the publisher's bonsai2 branch. The pinned card says stock upstream llama.cpp does not support PTQ1_0 correctly. The tested launch also used --spec-type draft-mtp --spec-draft-n-max 1, q4_0 K/V and reasoning effort medium.
The 8 GB result used the lean file and a card without a display for the 48K-context run. It is not proof that every 8 GB system can use that context or speed.
AMD Radeon support is labeled beta; the publisher had not reported a real-card AMD measurement at this revision. Do not transfer the NVIDIA results to Radeon.
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.
HumanEval+
95.12 / 100
Writing Python that passes extra edge-case tests. EvalPlus pass@1 on all 164 problems.
Benchmark efficiency: 15.996 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 thinking-mode benchmark table for the compressed Bonsai model; not scores copied from its Qwen parent. Custom Prism llama.cpp kernels are required. LiveCodeBench version and IFEval aggregation are unspecified. Table row: HumanEval+.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Following precise instructions. These cards report IFEval without consistently specifying prompt-level versus instruction-level scoring. Treat this view as approximate.
Benchmark efficiency: 15.355 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 thinking-mode benchmark table for the compressed Bonsai model; not scores copied from its Qwen parent. Custom Prism llama.cpp kernels are required. LiveCodeBench version and IFEval aggregation are unspecified. Table row: IFEval.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Solving grade-school math word problems. Published answer accuracy; known small subsets are separate.
Benchmark efficiency: 16.255 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 thinking-mode benchmark table for the compressed Bonsai model; not scores copied from its Qwen parent. Custom Prism llama.cpp kernels are required. LiveCodeBench version and IFEval aggregation are unspecified. Table row: GSM8K.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Competition math from AIME 2025. Published answer accuracy.
Benchmark efficiency: 15.975 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 thinking-mode benchmark table for the compressed Bonsai model; not scores copied from its Qwen parent. Custom Prism llama.cpp kernels are required. LiveCodeBench version and IFEval aggregation are unspecified. Table row: AIME25.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Competition math from AIME 2026. Published answer accuracy.
Benchmark efficiency: 16.115 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 thinking-mode benchmark table for the compressed Bonsai model; not scores copied from its Qwen parent. Custom Prism llama.cpp kernels are required. LiveCodeBench version and IFEval aggregation are unspecified. Table row: AIME26.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Basic Python programming with extra tests. EvalPlus pass@1 on the sanitized 378-problem set.
Benchmark efficiency: 13.969 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 thinking-mode benchmark table for the compressed Bonsai model; not scores copied from its Qwen parent. Custom Prism llama.cpp kernels are required. LiveCodeBench version and IFEval aggregation are unspecified. Table row: MBPP+.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Math reasoning over all 500 problems. Published exact answer accuracy.
Benchmark efficiency: 16.614 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 thinking-mode benchmark table for the compressed Bonsai model; not scores copied from its Qwen parent. Custom Prism llama.cpp kernels are required. LiveCodeBench version and IFEval aggregation are unspecified. Table row: MATH-500.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Revised MMLU knowledge test. Kept separate from original MMLU and MMLU-Pro.
Benchmark efficiency: 14.982 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 thinking-mode benchmark table for the compressed Bonsai model; not scores copied from its Qwen parent. Custom Prism llama.cpp kernels are required. LiveCodeBench version and IFEval aggregation are unspecified. Table row: MMLU-Redux.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Published coding pass rates without a verified benchmark version. Do not compare these directly with the v6 view.
Benchmark efficiency: 15.146 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 thinking-mode benchmark table for the compressed Bonsai model; not scores copied from its Qwen parent. Custom Prism llama.cpp kernels are required. LiveCodeBench version and IFEval aggregation are unspecified. Table row: LiveCodeBench.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Berkeley Function Calling Leaderboard version 3. Kept separate from version 4.
Benchmark efficiency: 12.599 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 thinking-mode benchmark table for the compressed Bonsai model; not scores copied from its Qwen parent. Custom Prism llama.cpp kernels are required. LiveCodeBench version and IFEval aggregation are unspecified. Table row: BFCL v3.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Instruction following: prompt-level loose accuracy. Separate from IFEval and other IFBench metrics.
Benchmark efficiency: 12.444 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 thinking-mode benchmark table for the compressed Bonsai model; not scores copied from its Qwen parent. Custom Prism llama.cpp kernels are required. LiveCodeBench version and IFEval aggregation are unspecified. Table row: IFBench (prompt-loose).
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
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 5.947 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 5.95 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.
Custom Prism llama.cpp kernels required; check backend support. The selected weights use 5.947 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.