The selected file is 47.45 GB (44.20 GiB), but that does not establish a memory requirement or compatible runtime. The independent converter measured 46.6 GB of RAM when loading it on CPU without memory mapping. That one result leaves too many system-specific variables to call every 64 GB computer compatible.
Measured Q4_K_M result
Converter-reported llama-bench results from 2026-10-07; not independently reproduced.
Setup
Prompt processing
Generation
CPU only; GPU hidden
83 ± 13 tok/s over 512 tokens
12.3 ± 1.4 tok/s over 128 tokens
RTX 5070 for attention/shared expert; routed experts in RAM; -ngl 99 --cpu-moe; q8_0 KV
723 ± 18 tok/s over 512 tokens
18.0 ± 0.3 tok/s over 128 tokens
Same hybrid setup after 32,768 tokens of context
696 ± 12 tok/s over 512 tokens
11.7 ± 1.1 tok/s over 128 tokens
Test machine: AMD Ryzen 7 7800X3D, 128 GB DDR5-6400 (2 × 64 GB), NVIDIA GeForce RTX 5070 12 GB, 8 CPU threads and flash attention on. The hybrid rows disabled CUDA graphs after the converter observed instability. Slower memory, another GPU, another backend or different context length can change the result.
Runtime and evidence limits
Patched runtime required: at the pinned revision, stock llama.cpp, Ollama and LM Studio do not load the kolibri1 architecture. The converter supplies a patch and prebuilt fork; support in later releases must be checked rather than assumed.
Conversion path: the independent conversion dequantizes Aleph Alpha's block-scaled FP8 weights to BF16 before making Q4_K_M. It is not endorsed by Aleph Alpha or llama.cpp.
Scores below are not Q4 tests: Aleph Alpha's published benchmark values describe the official FP8 source checkpoint. The converter explicitly says no benchmark suite was run on Q4_K_M and quantization can reduce accuracy.
Long context: the converter recalled three hidden facts at up to 253,646 tokens on the measured hybrid setup, but calls this a smoke test rather than a long-context benchmark. At the full 262,144-token test it reports 6.4 GB VRAM use with routed experts in system RAM.
Untested backends: Vulkan and Metal were not tested. CUDA was used for hybrid and long-context runs but was not compared with reference logits.
Compare 128 GB computers without assuming compatibility
The measured computer used 128 GB of ordinary system RAM plus a discrete 12 GB GPU. The products below are different 128 GB configurations and were not tested with Kolibri-1. Treat them as capacity and platform comparisons: confirm operating-system support, a working patched runtime, usable memory, memory bandwidth and return terms before buying.
As an Amazon Associate I earn from qualifying purchases. Each button uses the catalog's exact US ASIN and johndowie-20. Current price and stock are not verified here.
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.
MMLU-Pro
80 / 100
Knowledge and reasoning across many subjects. Published accuracy; different evaluation settings can change the result.
Benchmark efficiency: 1.686 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: Aleph Alpha-reported MMLU-Pro CoT English result for the official FP8 Kolibri checkpoint. The shared evaluation uses each model's documented context and sampling parameters, with Kolibri reasoning effort set to high.
Publisher-reported source-checkpoint result, not an independent test of the selected Q4_K_M conversion. The conversion publisher did not rerun the benchmark suite; quantization can change accuracy.
Hard science questions. This is a science reasoning test, not an overall intelligence score.
Benchmark efficiency: 1.776 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: Aleph Alpha-reported GPQA Diamond English result for the official FP8 Kolibri checkpoint. The shared evaluation uses each model's documented context and sampling parameters, with Kolibri reasoning effort set to high.
Publisher-reported source-checkpoint result, not an independent test of the selected Q4_K_M conversion. The conversion publisher did not rerun the benchmark suite; quantization can change accuracy.
Solving coding problems. Version 6 only; published pass rates can vary with generation settings.
Benchmark efficiency: 1.81 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: Aleph Alpha-reported LiveCodeBench v6 result for the official FP8 Kolibri checkpoint. The shared evaluation uses each model's documented context and sampling parameters, with Kolibri reasoning effort set to high.
Publisher-reported source-checkpoint result, not an independent test of the selected Q4_K_M conversion. The conversion publisher did not rerun the benchmark suite; quantization can change accuracy.
Competition math from AIME 2026. Published answer accuracy.
Benchmark efficiency: 2.023 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: Aleph Alpha-reported AIME 2026 English result for the official FP8 Kolibri checkpoint. The shared evaluation uses each model's documented context and sampling parameters, with Kolibri reasoning effort set to high.
Publisher-reported source-checkpoint result, not an independent test of the selected Q4_K_M conversion. The conversion publisher did not rerun the benchmark suite; quantization can change accuracy.
Fixing real software issues with tools. Agent setup and external tools affect these scores. File size covers model weights only.
Benchmark efficiency: 1.399 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: Aleph Alpha-reported SWE-bench Verified result for the official FP8 Kolibri checkpoint using the Harbor agent harness. The shared evaluation uses each model's documented context and sampling parameters, with Kolibri reasoning effort set to high.
Publisher-reported source-checkpoint result, not an independent test of the selected Q4_K_M conversion. Agent setup and tools affect this score; the conversion publisher did not rerun the benchmark suite.
Agentic terminal-use task success. Publisher-reported percentages are retained, but agent scaffold, tools, timeouts and reasoning budget can materially affect results.
Benchmark efficiency: 0.584 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: Aleph Alpha-reported TerminalBench 2.1 result for the official FP8 Kolibri checkpoint using the Harbor agent harness. The shared evaluation uses each model's documented context and sampling parameters, with Kolibri reasoning effort set to high.
Publisher-reported source-checkpoint result, not an independent test of the selected Q4_K_M conversion. Agent setup and tools affect this score; the conversion publisher did not rerun the benchmark suite.
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 47.454 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 47.45 GB file. Allow extra memory for context and software. Check the model’s runtime notes and your inference software before buying.
Listing evidence: 2026-09-24. 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.
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 47.454 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.