A 2B-parameter Qwen3.5 fine-tune, part of the Opus-Distil line in the reaperdoesntknow open-weight portfolio. Text-generation / reasoning model, trained with Unsloth + Hugging Face TRL.
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.
Parameters2.2B
Context1,048,576
Weights4.5 GB
License—
AccessOpen weights
Monthly Downloads3.5M
Runs On
What it takes to serve Kimi-K3-DSpark (2.2B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 4.5 GB | 5.4 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 2.2 GB | 2.7 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 1.1 GB | 1.3 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Sep 18, 2026.
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.
Model Card
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. 64 query heads / 16 KV heads, and blocksize=7 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…
Excerpt from the card by RadixArk.
Configuration
- Architecture
- DSparkDraftModel
- Context length (tokens)
- 1,048,576
- Layers
- 5
- Hidden size
- 7,168
- Feed-forward size
- 14,336
- Attention heads
- 64
- Key/value heads
- 16
- Head dimension
- 64
- Vocabulary size
- 163,840
- Model type
- qwen3
Identity and Version
- Repository
- RadixArk/Kimi-K3-DSpark
- Publisher
- RadixArk
- Task
- Text generation
- Modality
- Text
- Library
- transformers
- Parameters
- 2.2B parameters
- Languages
- Not stated by the source
- Revision
- 3c5bac301d9cf392706189d82ed947feca6c2f0f
- First published
- 2026-07-27
- Last updated
- 2026-08-16
Files and Weights
7 files, 4.5 GB in total. The weights are 1 file totalling 4.5 GB in safetensors.
Weights1 file · 4.5 GB
Configuration3 files · 26.8 KB
Documentation1 file · 3.9 KB
Repository2 files · 1.5 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model.safetensors | Weights | 4.5 GB | ecd746459b4a |
| config.json | Configuration | 1.4 KB | — |
| dflash.py | Configuration | 19.1 KB | — |
| dspark.py | Configuration | 6.3 KB | — |
| README.md | Documentation | 3.9 KB | — |
| .complete | Repository | 19 B | — |
| .gitattributes | Repository | 1.5 KB | — |
License and Download
- License
- Not stated by the source
- Access
- Open weights, no gate
- Download size
- 4.5 GB
Download from RadixArk
Released by RadixArk through its official repository on Hugging Face.
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 4.5 GB |
| 16-bit | 4.5 GB |
| 8-bit | 2.2 GB |
| 4-bit | 1.1 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Compare Kimi-K3-DSpark
Questions About Kimi-K3-DSpark
How much GPU memory does Kimi-K3-DSpark need?
About 5.4 GB at 16-bit and 1.3 GB at 4-bit: the weights (2.2B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run Kimi-K3-DSpark on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
What is Kimi-K3-DSpark's context length?
1,048,576 tokens, from the maximum position embeddings in its published configuration.
Similar Models
An English Qwen3.5 2B checkpoint associated with the Trendyol Cybersecurity Instruction Tuning Dataset and exported in Transformers / Safetensors format. This release is intended for research and local experimentation. The repository does not currently publish benchmark or safety-evaluation results, so the model should not be treated as a validated cybersecurity authority. The configuration identifies a Qwen3.5 conditional-generation architecture with text and vision components. Use a recent Transformers release that supports this architecture. Dependency and device behavior can vary across Transformers versions. Pin a tested environment for reproducible use. - Research on small-model…
local, typed decisions from qwen3.5-2b. one shared state is prefetched once, each question is isolated, every allowed answer is scored as a complete token sequence, and python returns validated choice, score, and noul results. this is an inference system, not rlcd training and not a clone of typesafe jev. it never calls typesafe. the published qwen checkpoint is unchanged; fp8 changes execution precision only. probabilities are normalized support over the candidates you provide, not calibrated correctness probabilities. unlike ordinary autoregressive json generation, the model does not write a response token by token. it scores only the values supplied by the caller. complete-sequence…
Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Built upon extensive training, Qwen3 delivers groundbreaking advancements in reasoning, instruction-following, agent capabilities, and multilingual support, with the following key features: - Uniquely support of seamless switching between thinking mode (for complex logical reasoning, math, and coding) and non-thinking mode (for efficient, general-purpose dialogue) within single model, ensuring optimal performance across various scenarios. - Significantly enhancement in its reasoning capabilities, surpassing previous QwQ (in thinking mode) and…
Topology-Aware Knowledge Distillation from Qwen3-30B-A3B → 1.7B TopologicalQwen is a 1.7B parameter model distilled from Qwen3-30B-A3B using Topological Knowledge Distillation (TKD) — a methodology that treats the teacher's output distribution over a concatenated token stream as a bounded variation (BV) function and decomposes knowledge transfer into three channels via the Mesh Fundamental Identity: 1. Smooth distillation (AC component) — Standard KL divergence over regions where the teacher's distribution varies continuously. This is what every other KD method does and stops at. 2. Jump corrections (D^j f) — Explicit correction terms at conceptual boundaries where the teacher's…
Extended Reasoning Distillation from Qwen3-30B-A3B-Thinking → 1.7B The most downloaded model in the Convergent Intelligence portfolio. Qwen3-1.7B-Thinking-Distil captures extended deliberation patterns from the Qwen3-30B-A3B Thinking teacher — the variant that generates long-form reasoning chains before committing to an answer — and compresses them into a 1.7B student via supervised fine-tuning on the longwriter-6k dataset. The Thinking teacher produces the richest signal of the three teacher variants in the DistilQwen family (Instruct, Thinking, Coder). Where Instruct distillation captures clean instruction-following and Coder captures hierarchical decomposition, Thinking distillation…