SAVRN
Search Contact SAVRN

SAVRN Model Hub · Comparisons

Kimi-K3-DSpark vs Llama-3.2-3B-Instruct

Kimi-K3-DSpark has 2.2B parameters and Llama-3.2-3B-Instruct has 3.2B parameters; at 16-bit, Kimi-K3-DSpark needs about 5.4 GB (1x MI300X from $1.85 an hour) and Llama-3.2-3B-Instruct about 7.7 GB (1x MI300X from $1.85 an hour).

Published metadata for 2 models, each read from its own repository.
Field Kimi-K3-DSpark
RadixArk/Kimi-K3-DSpark
Llama-3.2-3B-Instruct
meta-llama/Llama-3.2-3B-Instruct
Publisher RadixArk Meta Llama
Task Text generation Text generation
Modality Text Text
Parameters, as reported 2.2B parameters 3.2B parameters
Architecture DSparkDraftModel LlamaForCausalLM
Library transformers transformers
Context length 1,048,576 tokens Not stated
Repository size 4.5 GB 12.9 GB
Artifact formats safetensors safetensors, pytorch
License Not stated llama3.2
Access Open weights, no gate Access requested at publisher
Memory at 16-bit (weights and margin) 5.4 GB 7.7 GB
Cheapest GPUs at 16-bit, per hour 1x MI300X, $1.85 1x MI300X, $1.85
Memory at 4-bit (weights and margin) 1.3 GB 1.9 GB
Cheapest GPUs at 4-bit, per hour 1x MI300X, $1.85 1x MI300X, $1.85
Revision viewed 3c5bac301d9c 0cb88a4f764b
Downloads reported by the hub 3.5M 1.9M
Last observed 2026-09-18 2026-09-18

An evaluation row appears only where at least two of these models report the same benchmark with the same stated configuration, metric, unit and setup. Different evaluators stay named in each cell. Values are shown as reported: no unit conversion, no ranking.

SAVRN's Notes on Kimi-K3-DSpark

Read this as a companion, not a model you serve on its own. RadixArk trained Kimi-K3-DSpark as a DSpark speculator for the Kimi K3 target, a 2.2 billion parameter draft built for faster inference through speculative decoding. Five layers and 4.5 GB of weights keep it light: 5.4 GB of memory at 16-bit, 2.7 GB at 8-bit, 1.3 GB at 4-bit, and one MI300X with 192 GB at $1.85 an hour on-demand is the cheapest listing we track. That covers the draft only; the Kimi K3 target sizes your box.

The license field is empty, so get written terms from RadixArk before deployment; open access is not permission to run it commercially. Match your stack too: the checkpoint was trained with SpecForge on hidden states from a live SGLang target engine, and the 1,048,576 token context is the draft's ceiling, so the target needs the same window.

SAVRN's Notes on Llama-3.2-3B-Instruct

Memory is the easy part with this one. The 16-bit weights are 6.4 GB and it needs 7.7 GB to run, 3.9 GB at 8-bit, 1.9 GB at 4-bit, and our cheapest listed setup, one MI300X with 192 GB at $1.85 an hour on-demand, leaves most of the card free. That is the reason to pick 3.2 billion parameters tuned for multilingual dialogue, retrieval and summarization by agents: room for batches and other models.

The hard part is paperwork. Access is gated, so the publisher approves you first, and the license is listed as llama3.2 with no summary on record, so read the full terms before deploying commercially. Two more checks: the context length is missing from our record, and the repository is 12.9 GB across 16 files in both safetensors and pytorch formats against 6.4 GB of 16-bit weights, so fetch only the format you run.

Questions

Which is larger, Kimi-K3-DSpark or Llama-3.2-3B-Instruct?

Llama-3.2-3B-Instruct (3.2B parameters) is larger than Kimi-K3-DSpark (2.2B parameters), by the parameter counts their publishers report.

Which is cheaper to run, Kimi-K3-DSpark or Llama-3.2-3B-Instruct?

At 4-bit, Kimi-K3-DSpark fits on 1x MI300X from $1.85 an hour and Llama-3.2-3B-Instruct on 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Related Comparisons