<!-- Markdown mirror of https://getrigbase.com/models/raon-1-mini-preview/ . Source of truth is the HTML page. Data licence: CC BY 4.0, please cite RigBase (https://getrigbase.com/models/raon-1-mini-preview/). -->
Source: https://getrigbase.com/models/raon-1-mini-preview/
Last updated: 2026-11-18

[Models](https://getrigbase.com/models/) / Raon-1 mini Preview

# Best home rig for Raon-1 mini Preview

Quick answer

**How much VRAM does Raon-1 mini Preview need?**
The weights are about 17.6 GB at Q4_K_M (28.8B params × 4.89 bits ÷ 8, before context/KV cache). Add memory for context (KV cache) and the runtime. The smallest-memory setup with a measured result here, fully in memory, is Radeon RX 7900 XTX (24 GB) at 75.46 tok/s.

**What is the fastest measured hardware for Raon-1 mini Preview?**
RTX 5070 Ti: 220 tok/s generation (ik_llama.cpp-raon fork (CUDA 12.8), IQ3_KS experts / IQ4_KS_R16 down (13.85 GB, 3.86 bpw), q8_0 KV), from https://huggingface.co/McNiiby/Raon-1-mini-preview-IQ3_KS-GGUF.

Last updated: 18 Nov 2026

Current model 28.8B total · 4.3B active 5 measured results fastest plain decode: 220 tok/s  [model info ↗](https://huggingface.co/KRAFTON/Raon-1-mini-preview)

[Will it fit my GPU?](https://getrigbase.com/can-it-run/) [Best value GPU per euro →](https://getrigbase.com/guides/best-value-gpu-local-llm-europe/)

Real, measured generation speeds for Raon-1 mini Preview on home hardware, fastest first. Different quants/engines aren't directly comparable.

| Build | Quant | Engine | tok/s | Prompt processing (tok/s) | Cost | Source |
|---|---|---|---|---|---|---|
| [RTX 5070 Ti 16 GB: KRAFTON Raon-1 mini preview (28.8B-A4.3B) IQ3_KS at ~220-230 tok/s on an ik_llama.cpp fork, 128K context fits](https://getrigbase.com/builds/rtx-5070-ti-16gb-raon-1-mini-preview-iq3ks-ik-llamacpp-fork-mcniiby/) | IQ3_KS experts / IQ4_KS_R16 down (13.85 GB, 3.86 bpw), q8_0 KV | ik_llama.cpp-raon fork (CUDA 12.8) | **220** | 4,900 @4K · est. TTFT @8k ≈1.7 s* |  | [huggingface.co ↗](https://huggingface.co/McNiiby/Raon-1-mini-preview-IQ3_KS-GGUF) |
| [RTX 5070 Ti 16 GB: KRAFTON Raon-1 mini preview (28.8B-A4.3B) IQ3_KS at ~220-230 tok/s on an ik_llama.cpp fork, 128K context fits](https://getrigbase.com/builds/rtx-5070-ti-16gb-raon-1-mini-preview-iq3ks-ik-llamacpp-fork-mcniiby/) | IQ3_KS experts / IQ4_KS_R16 down (13.85 GB, 3.86 bpw), q8_0 KV | ik_llama.cpp-raon fork (CUDA 12.8) | **190** | – |  | [huggingface.co ↗](https://huggingface.co/McNiiby/Raon-1-mini-preview-IQ3_KS-GGUF) |
| [RX 7900 XTX (ROCm 7.2.1): KRAFTON Raon-1 mini preview IQ4_XS at 64.6 tok/s, 75.5 with opt-in fused kernels + backend sampler reuse](https://getrigbase.com/builds/rx-7900-xtx-raon-1-mini-preview-iq4xs-rocm-krafton-llamacpp-pr1/) | IQ4_XS GGUF | llama.cpp KRAFTON fork + PR #1 kernel profile (ROCm 7.2.1) | **75.46** | – |  | [github.com ↗](https://github.com/krafton-ai/llama.cpp/pull/1) |
| [RX 7900 XTX (ROCm 7.2.1): KRAFTON Raon-1 mini preview IQ4_XS at 64.6 tok/s, 75.5 with opt-in fused kernels + backend sampler reuse](https://getrigbase.com/builds/rx-7900-xtx-raon-1-mini-preview-iq4xs-rocm-krafton-llamacpp-pr1/) | IQ4_XS GGUF | llama.cpp KRAFTON fork + PR #1 kernel profile (ROCm 7.2.1) | **64.57** | – |  | [github.com ↗](https://github.com/krafton-ai/llama.cpp/pull/1) |
| [RTX 5060 Ti 16 GB: KRAFTON Raon-1 mini preview IQ4_XS with DFlash2 drafter at 42.1 tok/s, 256K window, 9 layers of experts in RAM](https://getrigbase.com/builds/rtx-5060-ti-16gb-raon-1-mini-preview-iq4xs-dflash2-krafton-llamacpp-niklaslenz/) | IQ4_XS GGUF (KRAFTON, 14.97 GiB) + DFlash2 Q8_0 drafter (699 MiB) | llama.cpp KRAFTON fork fb895662 (CUDA 13.3) | **42.1** ⚡ | 586 @8K · est. TTFT @8k ≈14.0 s* |  | [niklaslenz.com ↗](https://niklaslenz.com/v2/raon-1-mini-preview-5060ti-iq4xs-dflash2-llamacpp/) |

Prompt processing = prefill speed reported by the source, @ prompt length. * est. TTFT @8k is calculated from measured prefill, not measured. [Why it matters](https://getrigbase.com/can-it-run/#prefill).

## Quality cost of each quant for Raon-1 mini Preview

How far each quantization drifts from the full-precision (BF16) model, as published by the quant maker. Lower KL divergence (KLD) is closer to the original; same-top-token % is how often the quant picks the same next token as BF16. Q4_K_M is usually the sweet spot; below Q4 the drift grows fast, which is why RigBase only lists Q4-or-higher speed results.

| Quant | File size | Mean KLD vs BF16 | Same top token |
|---|---|---|---|
| IQ3_KS experts, trimmed head (released) below Q4 | 13.85 GB | **0.0931** | 86.97% |
| IQ3_KS experts, full head below Q4 | 13.96 GB | **0.0923** | 86.97% |
| IQ4_KS experts, full head (not released) | 16.26 GB | **0.0496** | 91.12% |

Source: [huggingface.co ↗](https://huggingface.co/McNiiby/Raon-1-mini-preview-IQ3_KS-GGUF). McNiiby (ik_llama.cpp quant publisher) card: KL divergence vs KRAFTON's Q4_K_M GGUF (19.7 GB, not BF16) on 40 x 512 tokens of held-out text; PPL ratio 1.040 / 1.040 / 1.001. These quant types load only on ik_llama.cpp forks. KLD from different publishers, datasets or llama.cpp builds isn't directly comparable.
Ad · this space is available **Sponsor this spot: reach people building local AI rigs.** One clearly-labeled slot per page, no tracking scripts.  [Sponsor this spot →](https://getrigbase.com/advertise/#inquiry)

**Rent before you buy:** try this model on a cloud GPU for a few dollars first: [Runpod](https://runpod.io?ref=77qcpl2c) (referral link; sign up with Google to get the new-user bonus)

## Memory math: how much VRAM / RAM does Raon-1 mini Preview need?

Raon-1 mini Preview has **28.8B parameters**, about **4.3B active** per token ([source](https://huggingface.co/KRAFTON/Raon-1-mini-preview): KRAFTON model card: 28.8B total / 4.3B activated MoE (6 of 128 routed experts + 1 shared), hybrid 18 Raon Delta Attention + 7 MLA layers, 262K context, Apache 2.0, trained from scratch; preview release (HF repo created 21 Oct 2026); HF API lists 28.78B params). Weight size = parameters × bits-per-weight ÷ 8, before context (KV cache) and runtime overhead.

| Format | Bits/weight | Weights | Smallest common memory size that holds the weights alone |
|---|---|---|---|
| Q8_0 | 8.50 | 30.6 GB | 32 GB |
| Q6_K | 6.56 | 23.6 GB | 24 GB |
| Q5_K_M | 5.70 | 20.5 GB | 24 GB |
| Q4_K_M | 4.89 | **17.6 GB** | 24 GB |
| Q3_K_M | 4.00 | 14.4 GB | 16 GB |
| Q2_K | 3.16 | 11.4 GB | 12 GB |
| As measured: IQ3_KS experts / IQ4_KS_R16 down (13.85 GB, 3.86 bpw), q8_0 | 3.86 | ≈13.9 GB | 16 GB |
| As measured: IQ3_KS experts / IQ4_KS_R16 down (13.85 GB, 3.86 bpw), q8_0 | 3.86 | ≈13.9 GB | 16 GB |

Bits-per-weight for GGUF quants are llama.cpp's own figures ([llama.cpp quantize README](https://github.com/ggml-org/llama.cpp/blob/master/tools/quantize/README.md)). "As measured" rows apply the source's stated bits-per-weight to the parameter count; real files differ slightly because some layers stay at higher precision. Try other combinations in the [VRAM calculator](https://getrigbase.com/can-it-run/).

**Why it can still run on smaller GPUs:** it is a mixture-of-experts model, so only about 4.3B of its 28.8B parameters are used for each generated token. Runtimes such as llama.cpp (--n-cpu-moe), ExLlamaV3 or Project Maya can keep most expert weights in system RAM and only the shared layers and hot experts in VRAM. Then you need **VRAM + system RAM ≥ weights**, and speed is limited mainly by RAM bandwidth.

**Measured on setups with less GPU memory than the Q4_K_M weights** (so part of the model lived in system RAM or was streamed from disk):

- [RTX 5070 Ti 16GB (16 GB)](https://getrigbase.com/builds/rtx-5070-ti-16gb-raon-1-mini-preview-iq3ks-ik-llamacpp-fork-mcniiby/): **220 tok/s** (IQ3_KS experts / IQ4_KS_R16 down (13.85 GB, 3.86 bpw), q8_0 KV)
- [RTX 5070 Ti 16GB (16 GB)](https://getrigbase.com/builds/rtx-5070-ti-16gb-raon-1-mini-preview-iq3ks-ik-llamacpp-fork-mcniiby/): **190 tok/s** (IQ3_KS experts / IQ4_KS_R16 down (13.85 GB, 3.86 bpw), q8_0 KV)
- [RTX 5060 Ti 16GB (16 GB)](https://getrigbase.com/builds/rtx-5060-ti-16gb-raon-1-mini-preview-iq4xs-dflash2-krafton-llamacpp-niklaslenz/): **42.1 tok/s** (IQ4_XS GGUF (KRAFTON, 14.97 GiB) + DFlash2 Q8_0 drafter (699 MiB))

## FAQ

### How much VRAM does Raon-1 mini Preview need?

The weights are about 17.6 GB at Q4_K_M (28.8B params × 4.89 bits ÷ 8, before context/KV cache). Add memory for context (KV cache) and the runtime. The smallest-memory setup with a measured result here, fully in memory, is Radeon RX 7900 XTX (24 GB) at 75.46 tok/s.

### What is the fastest measured hardware for Raon-1 mini Preview?

RTX 5070 Ti: 220 tok/s generation (ik_llama.cpp-raon fork (CUDA 12.8), IQ3_KS experts / IQ4_KS_R16 down (13.85 GB, 3.86 bpw), q8_0 KV), from https://huggingface.co/McNiiby/Raon-1-mini-preview-IQ3_KS-GGUF.

### Which GPUs have measured results for Raon-1 mini Preview?

RTX 5060 Ti 16GB, RTX 5070 Ti, Radeon RX 7900 XTX.

**Can it run on…** [RTX 5060 Ti 16GB](https://getrigbase.com/can-i-run/raon-1-mini-preview-on-rtx-5060-ti-16gb/) · [RTX 5070 Ti](https://getrigbase.com/can-i-run/raon-1-mini-preview-on-rtx-5070-ti/) · [Radeon RX 7900 XTX](https://getrigbase.com/can-i-run/raon-1-mini-preview-on-radeon-rx-7900-xtx/)

Related: [what GPU do you need for 2026 models](https://getrigbase.com/guides/what-gpu-for-qwen3-6-gemma-4-gpt-oss/) · [all builds](https://getrigbase.com/builds/)

**Search answer:** [How much VRAM do you need for Raon-1 mini Preview?](https://getrigbase.com/vram/raon-1-mini-preview/) · [Will it fit my GPU?](https://getrigbase.com/can-it-run/)
