Spark X2.5 Surpasses 2 Million Downloads, X2.6 Set to Go Open Source Soon

From 1 Million to 2 Million in Just 15 Days

Just 15 days after crossing the 1 million download mark, the open-source model Spark X2.5 has officially surpassed 2 million cumulative downloads (2,029,840). This growth comes almost equally from two parallel paths: official channels contributed 1,019,681 downloads, while derivative builds independently released by community developers contributed 1,010,159—nearly a 50/50 split.

Official channels span 6 platforms and 48 model repositories, while 109 community projects have published independently. For a lightweight model, this "official + community" dual-driven download structure is worth noting: it shows the model isn't just being pulled through official distribution pipelines, but is also being extensively modified, quantized, and re-released.

Official Channel Data Distribution

Downloads across the 6 official platforms show an extremely concentrated pattern.

Hugging Face Dominates Overwhelmingly

Hugging Face contributed 985,975 cumulative downloads across 12 model repositories, with 165,323 new downloads in the past 7 days. By comparison, the combined cumulative downloads of the other 5 platforms total less than 34,000. Ollama stands at 15,851, ModelScope at 11,116, Modelers at 5,210, AtomGit at 1,513, and SCNet at just 16.

Official channels added a total of 167,689 downloads in the past 7 days. This means all official platforms outside of Hugging Face added fewer than 2,400 downloads in a week.

Two Sides of Distribution Concentration

A single platform carrying nearly all official traffic binds the model's visibility and download behavior tightly to Hugging Face's ecosystem rules. For users, this reduces search costs; for the project team, it means distribution risk is relatively concentrated.

Hugging Face's ability to achieve this near-monopoly distribution position relates to its ecosystem role: it simultaneously serves as an entry point for model hosting, version management, model card documentation, inference endpoints, and dataset distribution, and offers direct integration with libraries like transformers and diffusers. Developers can complete their entire path from search to download to trial run within the same site, and model repository likes, download counts, and trends rankings in turn influence visibility, creating a positive feedback loop. In contrast, platforms like Ollama, ModelScope, Modelers, AtomGit, and SCNet each target different regions or usage habits—Ollama leans toward a one-click local-run command-line experience, ModelScope targets domestic Chinese users, and SCNet is tied to specific computing networks. Their smaller distribution volumes may reflect differences in user base, or may simply be repeated coverage after the same users have already downloaded from Hugging Face—something to keep in mind when interpreting each platform's numbers.

Community Builds Are the Other Half of the Story

The 109 community derivative projects generated a combined 1,010,159 downloads, nearly matching official channels. This figure reveals an easily overlooked fact: in local deployment scenarios, users tend to consume third-party pre-made quantized versions rather than downloading official original weights directly.

GGUF and MLX Are the Mainstream Formats

Among community builds, there are 45 GGUF builds, 13 MLX builds, 4 GGUF+imatrix builds, and several ONNX and edge-device variants. The top-downloaded projects are all GGUF format.

Build Project	Format	Downloads
peculiar-ragdoll/Sharp-Spark-X2.5-4B-GGUF	GGUF	435,736
abenzerps/Spark-X2.5-4B-GGUF	GGUF	378,904
SC117/Spark-X2.5-4B-abliterated-FIT-GGUF	GGUF	40,000
ngquocvinh/Spark-X2.5-4B-GGUF	GGUF	11,489
abenzerps/Spark-X2.5-4B-MLX-8bit	MLX	10,400

The top two community GGUF builds each surpassed 370,000 downloads, both exceeding any single source in official channels except the main Hugging Face repository. Note that these are all third-party builds rather than official releases—verify each model card carefully before use.

Abliterated (refusal-removed) variants also appear among community builds. Such modifications are not uncommon in the open-source community, but usage risks must be assessed independently.

GGUF is a model file format introduced by the llama.cpp project in 2023 to replace the earlier GGML. It packages model weights, tokenizer, prompt templates, and quantization metadata into a single file, with a forward-compatible versioning mechanism, making it the de facto universal exchange format for the local inference community. GGUF's key value lies in quantization: through different precision schemes like Q4_K_M, Q5_K_M, and Q8_0, the same model can be compressed to roughly one-quarter to one-half of its original size while maintaining usable quality on consumer CPUs, Apple Silicon, or low-VRAM GPUs. This is why the top-downloaded community builds are almost all GGUF—users can drag them directly into LM Studio, Ollama, or llama.cpp after downloading, without needing to convert weights themselves.

MLX, meanwhile, is an array computation framework Apple introduced for its own chips. MLX-format models are specifically optimized for the unified memory architecture of M-series chips, typically offering better inference speed and memory efficiency on Macs than general-purpose solutions. The parallel existence of GGUF and MLX reflects the fragmentation of local deployment scenarios: choose GGUF for cross-platform universality, MLX for performance within the Apple ecosystem.

imatrix refers to importance matrix, a quantization calibration method introduced by llama.cpp. Traditional quantization applies a uniform compression strategy to all layers, whereas imatrix first uses a small batch of calibration text to measure each weight's importance to the output, then allocates different quantization precisions to different tensors, concentrating error on unimportant small weights. In practice, imatrix quantization typically preserves more model capability than ordinary quantization at the same bit width, with the gap being particularly pronounced at low bit rates (such as Q2, Q3). A build with imatrix in the filename means the creator ran this extra calibration step during quantization, so the file size may be slightly larger and generation quality more stable, but it also depends more heavily on the representativeness of the calibration data.

As for abliterated (refusal-removed) variants, these identify and reduce direction vectors within the model related to refusal behavior at the weight level, so the model no longer outputs refusal responses. Such modifications have circulated in the open-source community for a long time, at the cost of weakening the model's safety alignment—output may contain harmful or inappropriate content—and different implementation methods vary widely, so compliance and risks must be assessed independently before use.

Spark-X2.6 Preview: Dual Sizes of 4B and 1.7B

The official announcement confirmed that Spark-X2.6-4B and Spark-X2.6-1.7B will go open source soon, with specific upgrade details to be revealed at release.

Continuing the Small-Size Route

X2.6's two sizes follow directly from X2.5's product line—both 1.7B and 4B are scales that can run on consumer hardware. The team expressed hope to "explore the possibilities of local AI together with everyone," with wording still pointing to edge-device and personal deployment scenarios.

Community Feedback Becomes a Pre-Release Variable

The team simultaneously opened the XHToken community repository to collect improvement suggestions and directs followers to the @SparkLLM X account for release updates. Actively soliciting needs before the official release shows the team treats the community build ecosystem as part of product iteration, not just the end of the distribution chain.

How to Use Spark X2.5

Spark X2.5 currently offers two sizes, 1.7B and 4B, with variants including Base, GGUF, FP8, and INT8.

• A complete collection is available on Hugging Face, and the team also notes that liking it helps more developers discover the model
• Individual repositories are Spark-X2.5-1.7B and Spark-X2.5-4B
• Ollama also provides a model entry point for SparkLLM

Data Caveats to Keep in Mind

The team explicitly states in footnotes that all numbers are cumulative all-time downloads as of October 14, 2026, and that platforms do not use a unified definition of "one download"—some count by file, others by repository. Modelers provides owner-level aggregate data, and SCNet's 7-day data shows almost no change.

More importantly, download counts do not equal unique user numbers. The same user may download multiple times across different platforms and different builds, so the 2 million total is better treated as a popularity indicator rather than a user-scale measure. For assessing actual model adoption, the structure of community builds and official channels each accounting for half may be more analytically valuable than the total itself.

For developers currently choosing a local model, one signal worth considering is this: the community has already produced a substantial volume of quantized artifacts around X2.5, meaning adaptation paths for mainstream inference frameworks and consumer hardware have likely already been paved.