Model reference · open weights
Kimi-K3-DSpark is an open-weight language model from RadixArk, listed in the AxForge catalogue. AxForge can bring it up on EU-owned hardware for you on request — with the licence handled where one is required.
About
Kimi K3 DSpark speculator Overview A long-context DSpark speculator for Kimi K3. It supports context lengths of up to 1 million tokens. A DSpark speculator for the Kimi K3 target, enabling faster inference through speculative decoding. DSpark extends the DFlash parallel-draft backbone with a Markov logit-bias head and a per-position confidence head. This checkpoint was trained with SpecForge using hidden states from a live SGLang target engine. Model Specifications - Base model: moonshotai/Kimi-K3 - Format: Safetensors (single-file BF16, 2,249,289,601 parameters) - Draft: 5 full-attention Qwen3-style GQA layers, hidden size 7168, 64 query heads / 16 KV heads, and blocksize=7 - Verification width: 1 current token + 7 draft tokens - Auxiliary target layers: [7, 23, 51, 67, 83] - Trained context: 65,536 tokens - Target weights: embedding and unembedding weights are not included Evaluation Results acclen is SGLang's histogram-native request acceptance length, averaged within each question and then equally across questions. RULER V2 uses the 1M input configuration. Actual prompts span 1,000,432–1,047,925 tokens; partition acclen is 4.4658 for MK, 4.3081 for MV, and 3.9919 for QA. AIME26 acclen by output length Serving with SGLang SGLang Cookbook provides Kimi K3 deployment recipes. YaRN-16 is enabled in the published draft config by default with originalmaxpositionembeddings=65536 and maxpositionembeddings=1048576; no separate draft config override is required. Training Details - Framework: SpecForge online distillation, with hidden states captured from a frozen Kimi K3 target served by a live SGLang engine. Draft trained from random initialization. - Loss: 0.1 CE + 0.9 L1 distillation + 1.0 confidence BCE, decay gamma 4.0, with 512 sampled anchors per sequence and blocksize=7. - Topology: 4 nodes × 4 GB300 (16 ranks) — 2 × TP8 target replicas, DP2 sampler, FSDP16 SHARDGRADOP on the draft, TP-batch scatter. Batch 8 per replica × 32 accumulation steps × 2 replicas = global batch 512.
Summarised from the published model card. Read the full card on the HuggingFace links below.
Specifications
| Maker | RadixArk |
|---|---|
| Type | Language models |
| Parameters (lead) | 2.2B |
| Context | 1024k tokens |
| Variants | 1 |
| Runs with | transformers |
| Released | 2026-07-27 |
| Popularity | 3.3M downloads / month |
| Likes | 54 |
| Licence | Commercial licence needed |
How it works
Variants
Open weights ship in several sizes and precisions. One page, all the variants — pick the one that fits your GPU. VRAM figures are estimates from model size.
| Variant | Params | Precision | VRAM | Fits 16 GB | Weights |
|---|---|---|---|---|---|
| Kimi-K3-DSpark | 2.2B | BF16 | ~5.2 GB | ✓ | Weights ↗ |
Using it via the API
Once AxForge deploys kimi-k3-dspark for you, it answers on the OpenAI-compatible API — the same base URL and keys as every other model. (kimi-k3-dspark below is illustrative; you get the exact model name on deployment.)
$ curl -sS https://api.axforge.ai/v1/chat/completions \
-H "Authorization: Bearer $AXFORGE_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"kimi-k3-dspark","messages":[{"role":"user","content":"Hello"}]}'
Details
Tags
Licence
The weights are open but its licence needs a commercial agreement for business use. AxForge can arrange that licence and host the model for you — you pay AxForge, we settle with the model’s maker. Ask us for a quote. Read the licence ↗