Open Source
Agens Engine is open source: Volundr's context window in a 16 GiB GPU budget
Why we wrote our own inference engine for Agens Volundr, where the memory goes, what it measures on a 16 GiB budget, and its limits. Version 1.3.5 is out.
Insight into the future
and lead change
Engineering notes, product updates and industry perspectives from the team behind HashAhead, Claway Team, Nomogram AI and AidVengers.
Technology Blog
Model Release
Agens Volundr 32B v1: long agent sessions on hardware you own
Only 18 of 72 layers keep a KV cache. What is inside, what v1 fixes over the Preview, the builds, where the model comes from and its known limitations.
HashAhead Engineering
Chain Group: how HashAhead reaches sub-second finality
An engineering walkthrough of the Chain Group architecture that anchors HashAhead — horizontal scaling, PoS consensus tuning, and the trade-offs we accept to keep block confirmation under one second.
Nomogram · Medical AI
Compliance-first medical AI: lessons from Nomogram for SMILE
What it takes to ship AI inside an ophthalmic clinic — OD/OS dual-eye analysis, role-based audit logs, PDF reporting and how Nomogram keeps surgeon data isolated end-to-end.
Industry Trend Analysis
Web3 × AI
The blockchain × AI convergence — what we see from the trenches
Notes from building HashAhead and the Claway / Nomogram AI stack side by side. Where the two technologies amplify each other, where they collide, and the architectures we keep coming back to.
AI Workforce
From chatbots to AI employees: the agent-platform shift
Why modular AI agents are replacing single-purpose chatbots in customer service, knowledge management and process automation — and what Claway Team's deployments taught us about scope.
Social Impact
On-chain transparency for community impact
How AidVengers uses tamper-proof donation tracking and on-chain supply distribution to keep disaster-relief workflows accountable — design choices and open questions.
Information Updates
Open Source
Agens Engine 1.3.5: steadier under long-context load
Long agent sessions hold up under load. With 48 concurrent ~30K-token sessions, failed requests went from 14 to 0 and the median time to first token from 25.5 s to 1.1 s (two 141 GB GPUs, compute capability 9.0, --tp 2, Volundr INT4, server mode, MTP). Agens Engine is open source under MIT.
Product
Lawway 律道: AI legal workflows that run inside your firm
A platform for Hong Kong solicitors and barristers. The model and the data stay inside the firm; the AI drafts, and the lawyer decides. Demos for invited firms.
Model Release
Agens Volundr 32B v1 and Agens Engine are out
Volundr v1 is an open model of about 32B parameters in which only 18 of 72 layers keep a KV cache, under Apache-2.0 on Hugging Face and ModelScope. Alongside it we open-sourced Agens Engine (MIT), the inference engine that runs its 16 GB build within a 16 GiB GPU budget.
Product Launch
Meet Blockway AI — a Cantonese-first AI assistant for iPhone, Android and the web
Blockway AI understands Cantonese the way Hong Kong people actually write it, along with English and Mandarin — mixed in one sentence if that is how you think. Send it a photo or a document, get fresh answers from the web with sources, have answers read aloud in natural Cantonese, or speak instead of typing. Every account gets a free monthly allowance, with Plus from HK$38 a month; no ads and no tracking, and your account can be deleted at any time. The iPhone and iPad app is in App Review and Android follows; the web version is open today with the same account.
Product Launch
Eduway is now open — an AI workbench built for Hong Kong teachers
Lesson prep, classrooms, materials, homework and school admin in one place, in 廣東話 and against the EDB curriculum. An AI toolbox drafts a worksheet in about five seconds; homework gets a first pass from AI before the teacher reviews it; and the Cantonese dictation cycle runs itself, re-testing wrong characters on days 1, 3 and 7. One rule holds the product together: nothing reaches a student until a teacher has read it. Free to register with a school email, with a demo that needs no account.
Product Launch
Codeway is now open — a coding agent for your terminal, and a chat on the web
Blockway's coding-agent harness for Agens Pilot is available today. In the terminal it reads your repository, runs commands in a sandbox, and edits and verifies; on the web it is a chat that sees and makes images and reads the web. One account, one monthly allowance, both places. Install with npm install -g codeway.
Product Launch
Introducing Agens Pilot — Blockway's Cantonese-first open model, built on Qwen3.8-27B
A post-trained fine-tune of Qwen3.8-27B that answers in 廣東話, gives balanced, multi-perspective answers on complex questions (85% vs 60% for the base, independently judged) with universal safety kept, and powers the Claway and Codeway agent harnesses. Three official builds (BF16 / FP8 / INT4) on Hugging Face under Apache-2.0.
Product Launch
AidVengers goes live — a blockchain platform for disaster relief and community support
AidVengers launches as a standing community-service initiative, bringing on-chain donation tracking, supplies-distribution management and a verifiable bridge between affected communities and their helpers.
Product Update
Blockway expands the AI portfolio: Nomogram for SMILE, plus image and audio/video AI
Blockway broadens its AI line-up with medical-grade Nomogram analysis for SMILE refractive surgery and a new generation of image and audio/video creation tools, built on the group's shared blockchain and compliance stack.
Engineering
HashAhead Chain Group hits a sustained sub-second confirmation milestone
The HashAhead public chain crosses a stability milestone: sustained sub-second block confirmations on production validators, powered by the Chain Group architecture and PoS consensus refinements.