🚀 iFLYTEK Open Source · Bi-Weekly Report (Sep 21 – Oct 4, 2026)

[FenjuFu]

🚀 iFLYTEK Open Source · Bi-Weekly Report (Sep 21 – Oct 4, 2026)

---

🌟 Highlights

• Spark-X2.5 passed one million downloads across all channels four weeks after open-sourcing. As of Sep 29 the total was 1,044,578: 662K from six official channels (Hugging Face, Ollama, ModelScope, Modelers, AtomGit and SCNet) and 383K from 99 community derivative projects.
• Eight repositories merged 42 PRs: SkillHub 15 · iFly-Skills 9 · Astron Agent 7 · Astron RPA 5 · AstronClaw Tutorial 2 · Domux 2 · MemFlywheel 1 · HarnessClaw 1.
• An Astron Agent bug reported in the group was fixed the same day. When account A published an agent to the agent marketplace, account B got "model does not exist" when using it. The report came in on Sep 22 and the fix was merged that day (#1680 → main #1681). The same batch also fixed failed workflow publishing (#1678) and missing code-node uids in imported workflows (#1689).
• Two n8n integrations. The Astron RPA Community Node lets self-hosted n8n start, wait for and cancel RPA executions over MCP, and n8n-nodes-iflytek exposes all 11 iFly-Skills as 25 n8n operations.
• SkillHub's enterprise identity work moved forward. Feishu and DingTalk are now public login providers (#877, #880), and the first organization control-plane slices are merged (#896, #897, #902). Three community issues filed on Sep 20 (a DeepSeek Harness install target, Mermaid rendering, and Suite Bundle import over plain HTTP) were all fixed and closed the next day.
• HER Hack-Astron 🚀 Astron Open Source · Bi-Weekly Report (Jun 29 – Jul 12, 2026) #6 (math reasoning) announced its champion on Sep 21. Preliminary entries for the Spark-X2.5 On-Device Model Innovation Challenge are due Oct 11.

👥 Top 10 Active Community Contributors

This ranking is based on message volume across 23 exported open-source discussion-group records. Automated Q&A accounts, welcome broadcasts and ads were excluded, display-name aliases were merged, and nicknames were anonymized. Two people tied for 10th place and both are shown. Thank you to everyone who answered questions, asked questions, and built alongside us 🙌

---

📦 Release & Development Highlights

✨ Spark-X2.5: filling the packaging gaps after one million downloads

Downloads and interest (Hugging Face 30-day downloads, Oct 5 snapshot):

• Spark-X2.5-4B-GGUF had about 479K downloads (399K in the Sep 28 snapshot used in the last report), Spark-X2.5-1.7B-GGUF about 213K (up from 155K), and Spark-X2.5-4B about 38K.
• Spark-X2.5-4B reached 1,380 likes, and the Spark-X2.5 GitHub repository has 633 stars.
• In the Sep 29 count, community derivatives accounted for 36.7% of downloads; the most downloaded was abenzerps/Spark-X2.5-4B-GGUF (about 215K). Counting by base_model on Oct 3, the four official repos have 105 direct community children: 48 GGUF, 19 MLX, 9 in other quantization formats, and 29 fine-tunes or adapters (#14).

Packaging gaps (each tracked in an issue; data and pull requests are welcome):

• 🦙 Ollama (#13): tool calling works, but SparkLLM/Spark-X2.5-4B and -1.7B only have a BF16 :latest tag (8.2 GB for the 4B), even though Q4_K_M (2.6 GB) and Q8_0 builds already exist on Hugging Face. The model pages on ollama.com also don't show the tools / thinking capabilities, and the minimum version is given as v0.34.1 in the READMEs but v0.34.2 in a maintainer reply.
• 🍎 MLX and INT8/FP8 (#14): upstream mlx-lm already supports spark2_5, but there is no official MLX build yet. The official INT8 / FP8 repos are also missing from the Quantizations tree on the Hugging Face model pages, so they are hard to find.

Feedback from the Hugging Face community (discussions opened this cycle):

• 🧩 A community developer suggested chat template improvements in 4B #24 and published Sharp-Spark-X2.5-4B-GGUF, tuned for agentic coding. The team replied that it will review the changes and take them into account.
• 🐛 A team fine-tuning the 4B for Greek found that greedy decoding can emit an orphan UTF-8 byte token (128486). This produces invalid UTF-8 and makes llama-server return HTTP 500. The report includes a minimal repro and a mitigation (4B #25).
• 🔁 With Q4_K_M, a 128K context and thinking enabled, the model got stuck in a "semantic loop" on a coding task and wasted 8K–10K tokens (4B #26). This matches the thinking loops reported in the groups over the previous two cycles.
• ⚙️ The stock vllm/vllm-openai:latest image can't load the model (4B #23). vLLM needs the Spark-plugin; see the FAQ. The team is also looking for a way to submit the models to the Artificial Analysis leaderboard (4B #22).

---

💬 Community Q&A (selected)

These topics are distilled from anonymized WeCom open-source group records for Sep 21 – Oct 4. Answers from the automated Q&A bot are for reference only; maintainers' conclusions take precedence. 🟢 Resolved · ℹ️ Clarified · 🟡 Needs follow-up.

Astron Agent

• 🟢 An agent published by account A shows "model does not exist" for account B. The model was configured under Model Management; once the agent was published to the agent marketplace, other accounts couldn't use it. A maintainer confirmed the issue on Sep 22 and merged a fix the same day (#1680 / #1681), covering the hub and front services. The user first only recreated the containers and still saw the error; it was fixed once they pulled the latest images for both services again.
• ℹ️ Does the open-source edition have team spaces? Yes. Click the … next to "Personal Space" in the sidebar to add a space or open space management.

• 🟢 How do I enable the E2B sandbox for an agent that uses Skills? Go to Resource Management → Script Sandbox and open the E2B Sandbox dialog. Turn on the enable switch at the top of the dialog first, then enter your E2B API key and save. The user found the switch easy to miss, and a maintainer said it will be made more visible. When E2B is off, workflow code nodes fall back to the built-in isolated LangChain/Pyodide executor.

• ℹ️ Can a Skill script read environment variables from .env? No. Skill scripts run inside the E2B sandbox and can't see variables in the host's .env. Inject the secrets a script needs on the sandbox side, or pass them in as call parameters.

• ℹ️ Should the RPA platform under Resource Management switch to iFlytek Cowork? No. The open-source edition integrates with Astron RPA as documented; the Cowork migration notice only applies to the Astron RPA SaaS edition.

• 🟡 How do I configure authentication when adding an MCP server? The Q&A bot only gave a speculative answer, and no maintainer has followed up yet. Please describe your scenario in issues and we will add documentation.
• 🟡 Do open-source workflows support multimodal input and output? The Q&A bot didn't give a useful answer, and the user later said they had found the setting in the UI. This capability still needs to be described in the official docs.
• ℹ️ Is there an edition for China's domestic IT stack (Xinchuang)? There is no separate edition today. The ongoing Hack-Astron x openKylin Welcome to the iFLYTEK Open Source Community 👋 #1 challenge is about deploying Astron Agent and SkillHub locally on openKylin 2.0+, so keep an eye on its results and deployment write-ups.

Astron RPA

• ℹ️ The Astron RPA SaaS edition is becoming iFlytek Cowork. Does this affect the open-source edition? No. According to the official notice, the SaaS edition has been upgraded to iFlytek Cowork, the old SaaS edition will receive no more updates, and SaaS users should migrate by Oct 15, 2026. The open-source edition of Astron RPA will continue to be maintained and updated. Cowork is not open source at the moment; for commercial licensing questions, please contact the team through the Astron RPA website.

• 🟡 After deploying the official v1.1.6 images, the client shows "service error" on launch, although registration and login work. The user suspects a server/client version mismatch. Support has confirmed the user runs the officially packaged images and is investigating. Note that the Q&A bot's advice at the time, "upgrade to v1.1.5+", was wrong: v1.1.6 is the latest release. The bot issue has been logged for a fix.
• 🟡 The client won't install on Windows 10 IoT Enterprise LTSC 21H2; Windows reports "This app can't run on your PC". This has been passed to support. If you hit the same error on LTSC or other stripped-down editions, please open an issue with your OS version and installer details.

• ℹ️ How does it compare with commercial RPA products? Group members said it handles simple desktop and browser actions well, and suggested downloading it and judging it against your own use case.

SkillHub

• ℹ️ SkillHub's architecture docs on zread.ai contain garbled text. zread.ai is a third-party site that generates docs from repository code automatically. The authoritative sources are the official docs at iflytek.github.io/skillhub and the design docs in the repository.

Spark-X2.5

• 🟢 Does it run on Ollama now? Yes. Update Ollama to v0.34.1+ and run ollama run SparkLLM/Spark-X2.5-4B (or -1.7B). Ollama currently only has the BF16 build (about 8.2 GB for the 4B); if you are short on VRAM, use the Q4_K_M GGUF from Hugging Face instead (#13).
• 🟡 vLLM parallel inference reports an unsupported architecture. vLLM needs the Spark-plugin installed; the stock vllm/vllm-openai image can't load the model (see also 4B #23). Group members also pointed out that the models are small enough to run several llama.cpp instances side by side. If it still fails with the plugin installed, please open an issue with your logs.
• ℹ️ All model links in one place: 4B GGUF · the full collection (including FP8 / INT8) · ModelScope · Ollama SparkLLM/Spark-X2.5-4B · PocketPal on mobile.
• ℹ️ A community member is building a strategy workflow that handles more than 1M tokens of context on a CPU machine to test how well Spark-X2.5 holds up over long contexts. We look forward to the results.

---

🔭 Community Spotlight

HER Hack-Astron #6 · Spark-X2.5 Math Reasoning Challenge (results announced Sep 21)

This edition had a single award. 🏆 The champion is Spark-X2.5-4B Discussion #19 (HF @Forrest20231206 / GitHub @MittaPei). It ran the official 4B BF16 weights on vLLM over 60 stratified GSM-Symbolic items, once with a direct-answer strategy and once with native function calling plus a calculator. The tool arm scored lower (58.33% vs 75.00%), yet all 63 calculator calls succeeded; instead, 18 trajectories used up their two calls and then wrote out a third call as text rather than giving a final answer. All 120 raw trajectories can be rescored offline, the statistics use a per-configuration exact McNemar test with Holm correction, and the entry does not overclaim. It won on methodology. The full judging notes are in Issue #9.

📌 Featured entry (shared in the groups on Oct 4): add 1 to the right-hand side and a congruence goes from solvable to unsolvable. Can Spark-X2.5-1.7B tell the difference? 1.7B Discussion #32 (by zhengjiyue) tested 16 pairs of linear congruences with the original BF16 weights on an Apple M5 Pro with MLX, once with thinking off and once with it on: 64 real inference runs in total. Strictly complete solution sets: 17/32 with thinking off and 27/32 with thinking on, at the cost of about 5× the median generated tokens and about 6× the total time. Typical mistakes included saying 3 divides 31 despite noting gcd = 3, and computing 38 but leaving it out of the final answer.

Spark-X2.5 On-Device Model Innovation Challenge (preliminary entries due Oct 11)

Build an on-device application on the open-source Spark-X2.5 models. The total prize pool is ¥100,000, with ¥25,000 for first place and prizes for the top 10, plus fast-track internship and job offers with the Spark foundation-model team. Preliminary entries close in one week. 👉 Details and registration

Hack-Astron x openKylin #1 (judging)

The challenge ended on Sep 24 and judging ran from Sep 25 to 30. Results will be announced in Hack-Astron x openKylin #1.

Milestones and community news

• 💯 Sep 29: four weeks after open-sourcing, Spark-X2.5 passed one million downloads across all channels.
• 📊 Since Sep 20: the team has been asking the Hugging Face community how to get the 1.7B and 4B models evaluated for the Artificial Analysis leaderboard (4B #22). If you know a contact, please leave a comment.

---

🙌 Build With Us

Good places to start this cycle:

• Add Q4_K_M / Q8_0 tags and capability labels for Spark-X2.5 on Ollama, or write up a conversion and validation plan for an official MLX build (#13, #14).
• Reproduce the thinking-mode semantic loop with a minimal prompt (4B #26) and compare how different quantization levels and sampling settings behave.
• Use the astron-workflow-dsl Skill to generate workflows in a local agent, import them into Astron Agent, and report any that fail to import.
• Improve the Astron Agent docs: MCP authentication, multimodal support, and the difference between the latest and vX.Y.Z image channels (#1694).
• Try the Astron RPA Community Node and n8n-nodes-iflytek on self-hosted n8n, and contribute acceptance notes and example workflows.
• Pick up the SkillHub CLI suite command (#893), or join the discussion on new requests such as one-time Namespace invite links (#911) and custom Skill avatars (#910).
• Check whether the Astron RPA client installs on Windows LTSC and other stripped-down editions, and record the compatibility results.

Before you start, please leave a comment on the relevant issue so maintainers can confirm the scope and avoid duplicate work.

---

🔗 Explore the Astron Open Source Ecosystem

• Astron Agent — enterprise-grade agentic workflow platform
• SkillHub — self-hosted registry and governance for agent Skills
• Astron RPA — agent-ready RPA for desktop and web automation
• Dolphin MCP Pilot — production-ready MCP server for Apache DolphinScheduler
• Domux — low-latency smart-home command parsing model
• MemFlywheel — agent-native long-term memory
• iFly-Skills · AstronClaw Tutorial · HarnessClaw
• Spark-X2.5 and the XHToken organization · Spark-X2.5 on Hugging Face
• All repositories tagged iflytek-astron

Methodology: PR counts use GitHub merged_at timestamps for Sep 21 – Oct 4, 2026, across the 10 repositories tagged iflytek-astron and all public repositories in the XHToken organization (which merged no PRs this cycle). Astron Agent fixes merged first to bugfix/superteam and then to main are counted as separate PRs. Version information comes from GitHub Releases published_at dates and Hugging Face repository creation and modification times. Hugging Face downloads are the platform's 30-day figures and likes are cumulative, both as of Oct 5, 2026. The one-million all-channel figure is from Sep 29; platforms count a "download" differently, so the total is indicative only. The HER Hack-Astron #6 announcement and the closing of the three SkillHub community issues both happened on Sep 21, inside this window. Community questions are anonymized and summarized. All screenshots come from public posts in the open-source groups; images containing personal accounts, phone numbers, hostnames or organization details were excluded, and customer identities, group identities and environment-specific details are not disclosed.

Questions, ideas, or contributions to share? Join the community discussion, or open an Issue / PR in the relevant repository. Let's keep building open, production-ready infrastructure for enterprise AI agents 🚀