Lines of source code march down the terminal. You are the shredder at the bottom. Break every token before a line reaches you, fill the bucket, and the round resets faster and meaner.
On phones and tablets, on-screen left / right / boost pads appear automatically.
Runs entirely in your browser. High scores are saved to local storage on your own device — nothing is uploaded.
I built Token Breaker with Hermes Agent. Not “AI-assisted.” The agent did the heavy lifting on architecture, implementation, iteration, and tooling while I directed, reviewed, and steered. Everything ran locally — the codebase never left the machine.
DeepSeek-V4-Flash on a single ASUS Ascent GX10 (NVIDIA GB10, 128 GB unified memory). Not multi-Spark tensor parallel. Not vLLM.
Best one-box local coding setup I have found for this hardware. Vision lives elsewhere.
The GX10’s 128 GB of coherent unified memory is both the constraint and the opportunity. Higher-precision quants of DeepSeek-V4-Flash simply do not fit with usable KV cache and headroom. The asymmetric IQ2 imatrix quant from antirez keeps the dense and shared components — attention, router, shared expert, output head — at higher precision while aggressively quantizing the routed experts. That trade-off has proven surprisingly effective for tool-calling and code generation.
DSpark speculative decoding is the real multiplier. It is lossless relative to the target model and delivers a meaningful decode speedup without the quality cliff you sometimes get from aggressive drafting. Combined with Entrpi’s Blackwell-optimized CUDA path and its significant prefill and decode lifts over upstream ds4, the box stays responsive through long contexts and tool loops.
Hermes Agent sits on top of the OpenAI-compatible endpoint, so it gets the full tool surface, persistent memory, skill evolution, and the ability to work across large codebases without constantly fighting the context window. A 1M native context means the agent can usually hold current game state, recent changes, design notes, and the relevant prior code all at once.
Text-only is intentional. Vision models and multimodal pipelines live on other machines. Keeping the primary coding agent pure text reduces complexity and keeps the memory footprint predictable on a single GX10.
I still review everything, set direction, and make the final calls. This is not autonomous founder mode. It is a high-capability local coding partner that never sends the codebase off-machine. Token Breaker exists because that local loop is fast enough, private enough, and capable enough that building this way felt natural instead of forced.
For pure local coding agents on current single-GX10 hardware, this remains the strongest combination I have run: model capability, context length, tool fidelity, speculative speed, and full offline control.
Play Token Breaker. Break some tokens. Tell me what breaks — or what doesn’t.
The game and the stack that built it both live here. Tell me what you find, or read more on running AI locally.