XHToken · Released Oct 19, 2026 · observed 31 Oct 2026
Public-source support evidence
We ran the Q4_K_M GGUF on one Mac; see Hardware notes below. The architecture is listed in llama.cpp and MLX / mlx-lm.
XHToken/Spark-X2.6-1.7B
Spark-X2.6-1.7B is a text generation model published by XHToken on Hugging Face. We have not written a summary yet; see the model card ↗.
View source model ↗SUPPORT EVIDENCE
The model config architecture is listed in this runtime registry. This is family/architecture evidence only; exact checkpoint loading, model-specific configuration, hardware, and generation were not tested.
Checked 2026-10-31 01:03 UTC
Evidence source ↗The exact model and runtime-specific sources were checked; no positive model-level evidence was established.
Checked 2026-10-31 01:02 UTC
Version v0.40.13
Checked source ↗The exact model and runtime-specific sources were checked; no positive model-level evidence was established.
Checked 2026-10-31 01:02 UTC
Checked source ↗The model config architecture is listed in this runtime registry. This is family/architecture evidence only; exact checkpoint loading, model-specific configuration, hardware, and generation were not tested.
Checked 2026-10-31 01:03 UTC
Evidence source ↗Checked the exact model config and the runtime architecture registry; no matching declared architecture was found. This is unknown, not evidence of incompatibility.
Checked 2026-10-31 01:03 UTC
Checked source ↗MLX community uploads may target mlx-vlm; an upload observation does not establish mlx-lm compatibility.
MEMORY ESTIMATES
The summary estimate of about 4 GB is the recommended Q4_K_M quant (1.8 GB observed), from file size × 1.2 + 1.5 GB. Each card is the largest observed quant whose estimate fits that capacity.
Which file should I grab?
Q8_0 · 1.8 GB observed · about 4 GB ↗About 4 GB using file size × 1.2 + 1.5 GB. Fit is an estimate; compatibility, speed, context length and hardware have not been tested.
Which file should I grab?
Q8_0 · 1.8 GB observed · about 4 GB ↗About 4 GB using file size × 1.2 + 1.5 GB. Fit is an estimate; compatibility, speed, context length and hardware have not been tested.
Which file should I grab?
Q8_0 · 1.8 GB observed · about 4 GB ↗About 4 GB using file size × 1.2 + 1.5 GB. Fit is an estimate; compatibility, speed, context length and hardware have not been tested.
Which file should I grab?
Q8_0 · 1.8 GB observed · about 4 GB ↗About 4 GB using file size × 1.2 + 1.5 GB. Fit is an estimate; compatibility, speed, context length and hardware have not been tested.
Which file should I grab?
Q8_0 · 1.8 GB observed · about 4 GB ↗About 4 GB using file size × 1.2 + 1.5 GB. Fit is an estimate; compatibility, speed, context length and hardware have not been tested.
Which file should I grab?
Q8_0 · 1.8 GB observed · about 4 GB ↗About 4 GB using file size × 1.2 + 1.5 GB. Fit is an estimate; compatibility, speed, context length and hardware have not been tested.
Which file should I grab?
Q8_0 · 1.8 GB observed · about 4 GB ↗About 4 GB using file size × 1.2 + 1.5 GB. Fit is an estimate; compatibility, speed, context length and hardware have not been tested.
Which file should I grab?
Q8_0 · 1.8 GB observed · about 4 GB ↗About 4 GB using file size × 1.2 + 1.5 GB. Fit is an estimate; compatibility, speed, context length and hardware have not been tested.
Memory estimates are the observed file size plus fixed headroom (20% plus 1.5 GB runtime allowance); real needs also depend on context length, KV cache and runtime settings. Fit is an estimate; it does not establish runtime compatibility, speed, context length, or hardware performance. Text-weights-only estimates do not include a projector or total vision memory where unknown.
Terms: a quant is a lower-precision version of the weights; GGUF is a file format used by llama.cpp and compatible apps; MLX targets Apple silicon; vLLM is a model-serving engine for NVIDIA GPUs; the KV cache is stored attention state that grows with context length. Terms explained ↗
OBSERVED FILES
Sizes and file groups come from the listed uploader repositories. A size is marked as an estimated fit only for a complete observed GGUF quant group.
| Uploader | Format & quant | Observed size | Estimated fit | Files |
|---|---|---|---|---|
| XHTokenXHToken/Spark-X2.6-1.7B-GGUF ↗Observed 31 Oct 2026 | GGUF · Q4_K_M | 1.8 GB | 16 GB, 24 GB, 32 GB, 64 GB, 128 GB, 192 GB, 256 GB, 512 GBEstimate only; requires the complete GGUF quant group. | 1 file |
| XHTokenXHToken/Spark-X2.6-1.7B-GGUF ↗Observed 31 Oct 2026 | GGUF · Q8_0 | 1.8 GB | 16 GB, 24 GB, 32 GB, 64 GB, 128 GB, 192 GB, 256 GB, 512 GBEstimate only; requires the complete GGUF quant group. | 1 file |
| XHTokenXHToken/Spark-X2.6-1.7B-GGUF ↗Observed 31 Oct 2026 | GGUF · Quant unknown | 3.4 GB | UnknownUnknown: completeness or a complete GGUF quant group was not verified. | 1 file
|
CURATED TRACKING
No curated issues verified for this snapshot. An empty list does not mean there are no bugs.
HARDWARE NOTES
We ran the Q4_K_M GGUF (1.77 GB, XHToken/Spark-X2.6-1.7B-GGUF @cdc3d88) on an Apple M4 Max Mac, 128 GiB unified memory, macOS 27.0.1, Metal with llama.cpp b11912 (upstream, 961c376) on 28 Oct 2026.
Decode: 177.8 tok/s on a short prompt, 158.0 tok/s on a long prompt (medians of 3 runs).
Peak memory: 2.45 GiB (llama-server footprint).
llama-bench: pp512 3,988.7 tok/s, tg128 238.0 tok/s.
Answer checks: 2 of 4 prompts passed (quick pass/fail sanity checks, not a quality benchmark).
We ran XHToken's official Q4_K_M GGUF of this model.
Default chat template settings (no thinking toggle passed). The model wrote reasoning text on every prompt; decode speed counts every generated token.
Both failed checks hit the 512-token output cap while the model was still reasoning.
One machine, one run set: temperature 0, 8K context, 1 warm-up plus 3 runs per prompt. Short prompt is 78 tokens; long prompt is 1,539 tokens (each model counts with its own tokenizer). Weights were loaded from an external USB drive with mmap off; load time is not published for this run. Your speed will differ with hardware, context length and settings.
COMPARE
No comparison page is generated for this model yet. Browse all comparisons →
EMBED
Works in GitHub READMEs and Hugging Face model cards. The badge updates when the site rebuilds.