Spark-X2.5-4B and Spark-X2.5-1.7B: Compact On-Device Agent Models With a Native 1M Context

Spark-X2.5-4B and Spark-X2.5-1.7B went public on September 1, 2026, when SparkLLM — the XHToken (词元星火) subsidiary of iFlytek — open-sourced both compact models under Apache 2.0, with weights, code and deployment docs landing on Hugging Face and GitHub the same day. The headline spec is a native 1,000,000-token context window on models small enough to run on-device, backed by a claimed 200+ language vocabulary and a hybrid-attention design the vendor says is tuned for agentic work. What is knowable today from the repos is substantial; what is not yet confirmed is everything that only independent benchmarking can settle, because a model that shipped hours ago has no neutral evals yet. The X2.5 family also has a larger member coming — Spark-X2.5-293B, announced for September 7.

What the repos actually show

The Hugging Face collection carries six repositories: instruction-tuned Spark-X2.5-4B and Spark-X2.5-1.7B, their base checkpoints, and GGUF conversions of both. The 4B card describes a hybrid-attention architecture — one full-attention layer per three sliding-window attention layers — which is exactly the shape you want when the marketing number is 1M tokens, because full attention over a million tokens would otherwise be quadratically expensive. The card states a native context window up to 1M tokens, with a long-context training stage that extended sequence length to 1M during pretraining.

• Architecture — hybrid attention, one full-attention layer per three sliding-window layers; BF16 safetensors.

• Context — native up to 1,000,000 tokens; long-context training ran sequences to 1M.

• Languages — more than 200, per the model card (the 1.7B shares the claim).

• Pretraining — roughly 20 trillion tokens of web pages, books, academic publications, code and encyclopedic material, trained end to end on Huawei Ascend clusters.

• Post-training — SFT followed by large-scale RL across reasoning, coding, agentic behaviour and instruction following, with domain policies consolidated via a technique the paper calls MOPD.

• License — Apache 2.0 for both sizes, with no revenue or region clauses of the kind several other Chinese labs attach to open releases.

The agent numbers — and how much to trust them

The model card publishes a full agent-and-reasoning benchmark table, and it is uniformly vendor-reported and unreproduced — label it that way, because the interesting question for a day-old 4B is whether any of it survives contact with independent evals. On the vendor's own table, Spark-X2.5-4B scores 65.1 on BFCL-V4 (function calling), 75.1 on τ²-bench (long-horizon agent tasks), 40.9 on BrowseComp, 44.4 on SWE-Bench Pro, 90.7 on AIME 2026, 81.2 on HMMT February 2026, 74.2 on IMO-AnswerBench, and 67.4 on GPQA. The 1.7B gets its moment on a smart-home test: 90.3% end-to-end command accuracy at 0.85 seconds average latency on the Domux set. The vendor also claims the 4B "rivals cloud models 2–3× its size" on algorithm implementation, code completion and generation — a claim that is simultaneously the most useful sentence in the release and the one most in need of a neutral check.

What is structurally more interesting than any single number is that the release is aimed at agents from the start. The card lists integration with the Codex, Claude Code, OpenClaw and Hermes harnesses, tool-calling support with a dedicated parser in SGLang, and reasoning enabled by default (a runtime flag, enable_thinking, turns it off). In other words the vendor has positioned a 4B as a tool-use model, not a chatbot, which is a real bet: small-agent models are where the on-device market is actually moving.

Day-0 ecosystem, and the vLLM story

Spark-X2.5-4B and Spark-X2.5-1.7B are supported in vLLM from day one via an out-of-tree general plugin rather than an upstream merge. The integration lives in the XHToken/Spark-plugin repository: you clone it and uv pip install ., then serve the model with vllm serve and --trust-remote-code against a custom chat template. The SGLang path is a prebuilt Docker image launched with --tool-call-parser spark25 and --context-length 1048576 — a full million-token window in the launch command, which is the fastest way to sanity-check the context claim on real hardware.

Beyond vLLM and SGLang, the model runs on the llama.cpp fork, MLX (Apple silicon and Linux CPU), Ollama and LM Studio via GGUF, with fine-tuning supported through a LLaMA-Factory fork. Hardware support is the other headline: NVIDIA, Huawei, Hygon and HOUMO.AI — plus Apple silicon — which matters because it is rare for a Chinese-lab model to ship with Ascend, Hygon and domestic-chip deployment docs on day one, not as a promise.

What a 1M-token on-device model is for

The context length is the feature, and it points at specific jobs. A million tokens of native context on a 4B means an entire codebase, or a long document set, loaded into a model that runs locally — no RAG pipeline, no chunking. The vendor's own demonstration, built on its Loomy office tooling, has the 4B write a script to aggregate a sales spreadsheet, extract key metrics and trends, and generate a roughly 3,000-word bilingual report. The robotics angle is the other half: both sizes are positioned for on-robot and edge perception-and-execution, covering control, target tracking and navigation, which is a plausible niche for a 1.7B that answers in under a second.

The bigger play: Spark-X2.5-293B

The two small models are the opening move. On September 7, SparkLLM says it will release Spark-X2.5-293B, a 293-billion-parameter base model with, per the announcement, upgraded coding and agent capabilities. The X2.5 small models are effectively the on-device half of a two-tier story; the large model is where the family's ceiling gets set. Treating the 4B and 1.7B as the whole release would miss the direction of travel.

How to try it, and what to watch

The API is live on iFlytek's Xingchen MaaS platform and free for a limited time; the weights are on Hugging Face, ModelScope and the Ollama library. On the routing side, Spark-X2.5-4B is too new for any aggregator to be serving it yet — OrcaRouter does not route it today, and nothing on this page should be read as if it did. But the day a routed provider adds it, the economics change in a predictable way: OrcaRouter passes provider list price through with 0% markup, so whatever the vendor prices it at is what you pay, with the same API key you already use, and automatic failover so a day-0 model never has to be a production bet. That is the standard shape of how this blog sees a brand-new model, and it is worth saying plainly.

Four things will decide whether this release is a moment or a footnote. First, whether independent evals — an Artificial Analysis index entry, an arena placement, a reproduced τ²-bench run — come anywhere close to the vendor's table; a 4B scoring 90.7 on AIME is a big number, and big numbers from a release-day card are exactly the ones that get checked. Second, whether the 1M-token context holds up on the hybrid-attention stack at real serving loads, not just in the launch command. Third, whether the Ascend-trained weights behave on NVIDIA hardware in the field. Fourth, what Spark-X2.5-293B actually delivers on September 7. Until then, the honest summary of Spark-X2.5-4B and Spark-X2.5-1.7B is: promising, open, and entirely unverified — which, for a model that shipped this morning, is the correct status.