Spark X2.6 1.7B
XHToken's state-of-the-art 1.7B on-device reasoning and coding model. Features a 1M token context length, configurable thinking effort modes (none/low/high), and outsized performance on SWE-bench and autonomous planning.
The brutal truth
The undisputed champion of 1.7B on-device reasoning. Having a controllable thinking budget (none/low/high) plus a 1M token context under Apache 2.0 makes it the best lightweight agent brain available.
Files and Versions
Direct inspected tree artifacts from Hugging Face. Verified filenames, exact byte sizes, model weight shards, and configuration definitions.
Decision matrix
✓ RUN / USE THIS IF…
- •You want to ingest massive codebases or logs with up to 1M context tokens
- •You want controllable thinking modes (none/low/high) under Apache 2.0
- •Your 24GB GPU comfortably fits the model weights (~1.8GB VRAM) with dedicated headroom for inference.
- •You require 100% offline data privacy with zero third-party API logging, subscriptions, or prompt retention.
✕ WALK AWAY IF…
- •You need 70B-scale domain expertise across obscure topics
- •You cannot enable trust_remote_code in your Python environment
- •Your GPU has less than 1.8GB VRAM and you cannot tolerate slow CPU/system RAM offload paging.
- •You prefer a plug-and-play web app interface and don't want to manage local model checkpoints, drivers, or runtimes.
True cost of ownership
About
XHToken's state-of-the-art 1.7B on-device reasoning and coding model. Features a 1M token context length, configurable thinking effort modes (none/low/high), and outsized performance on SWE-bench and autonomous planning.
What it does
Download the weights from Hugging Face, pick a quantisation that matches your GPU, and run it locally with Ollama, llama.cpp, vLLM, or a compatible runtime. On-device thinking reasoning, 1M context & SWE-bench tool use runs on your own hardware — no per-token bill, no third-party data handling.
Key features
Massive 1M token context window in an ultra-compact 1.7B parameter footprint
Built-in controllable thinking budget (none/low/high) allows balancing latency vs deep reasoning
Punches way above its weight on SWE-bench tool use and multi-step planning under Apache 2.0
Use cases
Inline edits, refactors, and test generation inside the editor.
Flag null-safety, security, and style issues before merge.
Turn long procedural functions into typed, testable modules.
Generate integration tests from route definitions and schemas.
Quantization & Hardware Fit
VRAM Hardware Compatibility & Memory Allocation
| Quant | VRAM | Quality / Fit on 24GB | Install / Run Command (OLLAMA) | |
|---|---|---|---|---|
| Q4_K_M● ACTIVE | 1.8 GB | Standard · ✓ Fits (22.2GB spare) | ollama run xhtoken-spark-x26-17b | |
| Q8_0 | 2.5 GB | Standard · ✓ Fits (21.5GB spare) | ollama run xhtoken-spark-x26-17b | |
| BF16 | 3.6 GB | Standard · ✓ Fits (20.4GB spare) | ollama run xhtoken-spark-x26-17b |
Buyer doubts, cleared
Can I actually run Spark X2.6 1.7B on my 24GB GPU?
What's the fine print on the license?
Does the 1M tokens actually hold up in practice?
User reviews
Switched our internal workflow to Spark X2.6 1.7B after the last pricing change. Nobody noticed a quality difference, and the bill dropped by half.
The Q4_K_M build is the sweet spot. Anything larger needs serious hardware for marginal quality gains.
Solid for on-device thinking reasoning, 1m context & swe-bench tool use. The learning curve is real but the payoff is there. Start with the smaller build first.
The output reads naturally once you tune the prompt. It has cut my first-draft time roughly in half.
