Spark-X2.5: open 4B/1.7B agentic models that run on your laptop
Apache-2.0, 1M context, tuned to be driven by Claude Code and Codex — SWE-Bench Verified 41.6 in a 4B. Plus: Claude Code 2.1.273 adds LLM routing hints.
Copy markdown
What dropped: two Apache-2.0 agentic models
XHToken released Spark-X2.5 in 4B and 1.7B, both Apache-2.0 with a native 1M-token context and 200+ language coverage. It's a from-scratch small-model series built for on-device agents — not a distill of a bigger frontier model.
Built to be driven, not chatted with
It ships with adapters for Claude Code, Codex, OpenClaw and Hermes, and the agent scores back it up: BFCL-V4 65.1, MCP-Atlas 54.6, Workspace Bench 31.2 — trading blows with Qwen3.5-9B at under half the size.
Frontier-ish reasoning, in 4B
For a model this small the numbers are unusual: SWE-Bench Verified 41.6, SWE-Bench Pro 44.4, and AIME 2026 90.7. Enough that a laptop-class model can take a real bug-fix or math problem end to end.
How to run it today
GGUF quants are already up — Q4_K_M at 2.6GB, Q8_0 at 4.38GB — plus MLX, vLLM, Ollama and LM Studio support. The 4B has pulled roughly 167k downloads in a month, so the llama.cpp path is well-trodden.
The catch: read before you commit
One independent local test found it correct but 3–4x slower than bigger models on the same fix, and that 1M context is aspirational — KV cache hit ~41GB at just 65k tokens on a 48GB GPU. Multilingual output also broke down outside major European languages.
Build this weekend
A fully-local coding-and-tools agent: point Claude Code or Ollama at the Q4 weights and you get a laptop agent that calls MCP tools with zero API cost and no data leaving the machine — ideal for overnight batch runs where latency doesn't bite.
Elsewhere: Claude Code 2.1.273
The latest build adds gateway hint headers so you can steer LLM routing, plus remote-control session forking and fixes to MCP disconnects and auto-mode permission checks.