Blog
Insights, updates and articles on AI agent orchestration — from practical guides to deep dives on building with HelixML.
We Pointed Helix Cyber at OWASP Juice Shop
Oct 29, 2026A real Helix Cyber engagement, end to end: a gray-box web pentest against OWASP Juice Shop, run by open-weight models on self-hosted inference. Then the same job on two models, GLM-5.3-Flash and DeepSeek-V4.1-Flash: 67 and 58 independently verified findings, and each one caught serious bugs the other missed.
Read more →Serving Nanbeige4.3-33B to Coding Agents: NVFP4, a One-Line Template Fix, and Up to 65% More Agent Turns
Oct 28, 2026Nanbeige4.3-33B as a local model for Claude Code on vLLM: NVFP4 at 9,143 tok/s per RTX PRO 6000, a template fix for prefix caching, and 65% more agent turns.
Read more →9% faster GLM-5.3-Flash prefill with GPU-to-GPU IPC (and why TileLang didn't help)
Oct 26, 2026We profiled GLM-5.3-Flash on two RTX PRO 6000 GPUs, with the non-uniform memory access (NUMA) check and the SGLang settings that reproduce the result. A TileLang kernel ran 2x faster in isolation and changed nothing in production. A chunked PCIe all-reduce made prompt reading 9% faster, and a memory-placement bug explained a phantom 27% regression.
Read more →Optimizing Domyn Edge for Multi-Agent Sessions: Up to 61% More Turns
Oct 21, 2026Domyn Edge is a 4B model trained in Europe, with open weights, data and training code. We tuned it for many coding agents working at once: thinking off made tool calls reliable, FP8 doubled throughput, suffix decoding made code edits up to 8× faster, and Ramjet served 53-61% more agent turns than nginx on the same four GPUs. It also runs on a gaming RTX 4090 with the full 256K context.
Read more →Ramjet Now Mints API Tokens and Counts Every Request
Oct 20, 2026Ramjet, our open-source load balancer for vLLM and SGLang, can now issue its own API tokens and attribute every request to a token and a model. The first usage store needed 5.1 GiB of RAM for 1,000 active tokens. The SQLite version needs 7-25 MiB, and Parquet files measured about six times smaller again.
Read more →Basalt Runs Qwen3.8-Flash-Next at 229 Tokens/s on Two GPUs, One Request at a Time
Oct 15, 2026A 3.75-bit Basalt build of Qwen3.8-Flash-Next generated 229 output tokens per second for a single request on two RTX PRO 6000 GPUs, about three times our NVFP4 baseline, and fell behind it when four requests ran together. The configuration that worked, and the limits we measured.
Read more →Ramjet on DeepSeek-V4.1-Flash: Routing Still Pays When the Cache Never Fills
Oct 7, 2026On one 8×H200 server, DeepSeek-V4.1-Flash never filled more than 22% of its prompt cache. Sending each coding-agent session back to the engine that already read it still added 18% throughput. Here is what Ramjet did, and two ideas that did not work.
Read more →Attackers Have Uncensored LLMs. Your Security Models Refuse to Help.
Oct 4, 2026Refusal-free open-weight models are a working tool on one side of security: attackers download them without an account, while aligned models refuse 43.8% of fully defensive hardening prompts. What the asymmetry costs, measured.
Read more →Best Uncensored LLMs for Security Work: The 2026 Field Guide
Oct 4, 2026Which uncensored and abliterated models make sense for security work in 2026: the 27B abliterations, the GLM-5.3 and DeepSeek-V4.1 frontier, the quality costs each one publishes, and how to measure refusal on your own prompts before choosing.
Read more →Ramjet vs NVIDIA Dynamo: Which Router for Coding-Agent Traffic?
Sep 30, 2026On one 8×H200 server running GLM-5.3, Ramjet kept 92% of agent prompts in cache against 83-85% for NVIDIA Dynamo 1.5.0's cache-aware router, and completed 22-29% more agent turns a minute on the same engine. Dynamo is built for larger fleets, where we expect it to pull ahead.
Read more →Serving Full GLM-5.3 to 48 Coding Agents From One 8×H200 Server
Sep 29, 2026The full 753-billion-parameter GLM-5.3 fits on one 8×H200 server. Splitting its attention eight ways, routing each agent session back to its own GPU with Ramjet and adding a host-memory cache took the server from 30 to about 100 agent turns a minute, enough for 32 to 48 continuously working coding agents.
Read more →Meet Your Chief of Staff
Sep 28, 2026Join the Helix Cloud private beta and see a Chief of Staff bring in a specialist to tackle a real task.
Read more →Swift 1.5 Flash-Next Cut Qwen3.8's Thinking Tokens in Our Pilot
Sep 27, 2026We deployed UkisAI's Swift 1.5 Qwen3.8-Flash-Next on four RTX PRO 6000 GPUs and compared completed answers, thinking tokens, tool use and serving speed with our previous Qwen checkpoint.
Read more →GLM-5.3-Flash Ran Out of Cache Snapshots, Not Cache Tokens
Sep 26, 2026Our GLM-5.3-Flash replicas kept forgetting cached prompts while their 500,000-token KV cache sat mostly empty. On a hybrid linear-attention model the limit was 28 state snapshots. Here is how we found it and what fixed it.
Read more →How to Build Applications on Top of Helix
Sep 22, 2026Keep your application's workflow and use Helix to run complex agent tasks in isolated workspaces, report progress, and return useful outputs.
Read more →Self-Hosting Kev on an RTX PRO 6000 With Ramjet
Sep 22, 2026How we run Kev as an on-premises AI decision model on an RTX PRO 6000, with Ramjet routing TypeSafe System One and OpenAI APIs.
Read more →Running GLM-5.3-Flash on 2, 4 or 8 RTX PRO 6000 GPUs
Sep 14, 2026Our measured GLM-5.3-Flash TP2 recipe reaches 388 output tok/s and nearly 6,000 prefill tok/s on two RTX PRO 6000 GPUs, then scales by adding independent replicas.
Read more →DeepSeek V4.1 Flash: Why Its Encoder, Engram and KV Cache Matter
Sep 10, 2026DeepSeek V4.1 Flash splits prompt reading from generation, adds 196B Engram memory and compresses key-value cache. Here is what that means for serving.
Read more →Automating LinkedIn Recruiter When There Is No API
Sep 3, 2026LinkedIn Recruiter has no public API and punishes seat sharing, so the usual integration approach fails. How we ran overnight candidate sourcing for a recruitment agency by giving an agent a desktop and a browser.
Read more →Your agent shouldn't be you
Sep 1, 2026How each coding agent in Helix gets its own credential, its own network, and write access to two branches and nothing else. Including the parts we got wrong and what we still haven't solved.
Read more →Penetration Testing Cost in 2026: A Buyer’s Guide
Aug 31, 2026Most penetration tests cost $10,000–$30,000 in 2026. Compare prices, hidden costs, testing cadence, PTaaS, and continuous security testing for AI agents.
Read more →How to Run Claude Code Safely: From Permission Prompts to Enterprise Sandboxes
Aug 31, 2026Run Claude Code safely at any scale: permission modes, devcontainers, microVMs, and isolated agent desktops — a containment ladder for Claude Code security.
Read more →Security Data Lake vs SIEM: Stop Paying to Delete Your Own Evidence
Aug 31, 2026SIEM ingest pricing forces teams to drop the telemetry they need most. When a security data lake beats a SIEM, when it doesn't, and how the hybrid works.
Read more →Self-Hosted E2B Alternatives: 6 AI Agent Sandboxes Compared
Aug 31, 2026Daytona went closed source in June 2026 and self-hosting E2B is a real project. Six E2B alternatives and Daytona alternatives for AI agent sandboxes, compared.
Read more →One Qwen Model, Two Speeds: What Smart Routing Bought Us
Aug 28, 2026Ramjet routes short and long responses across two Qwen3.8 replicas while preserving prompt-cache locality, improving measured throughput by up to 79%.
Read more →GLM-5.3-Flash on RTX PRO 6000, Part 1: Getting Day-Zero Serving to Work
Aug 27, 2026We got the 321B-parameter GLM-5.3-Flash serving in NVFP4 on four RTX PRO 6000 GPUs. Here is what worked, what broke, and why it is still an experiment.
Read more →Qwen3.8-Flash-Next on Eight GPUs: What Actually Helped
Aug 27, 2026We ran Qwen's Qwen4 preview across eight RTX PRO 6000 GPUs. Speculative decoding helped at low load, but smarter routing cut repeat-request startup from 2.5 seconds to 287 ms.
Read more →A Better lm_head for Qwen3.8-27B: How We Tested and Shipped It
Aug 25, 2026A new BF16-head checkpoint promised better accuracy. We matched it against production, measured 7.5% faster decode with no regression in our agent corpus, then rolled it across eight GPUs one at a time.
Read more →We Doubled Our Inference Throughput by Reading a Log Line
Aug 23, 2026Our 8× RTX PRO 6000 box looked saturated at 4,256 tok/s. It wasn't — it was queueing behind a state cache we'd never configured. One flag took c192 from 3,833 to 7,882 tok/s, and gating it honestly meant rebuilding our test harness.
Read more →Chasing a 454 tok/s tweet: a day of tuning Qwen3.8-27B on the RTX PRO 6000
Aug 22, 2026We tried to reproduce a viral single-GPU benchmark on our production box. We got 335, found out why the last 120 tok/s is physically locked out on Server Edition cards, and walked away with a +12% win we actually shipped.
Read more →SGLang vs DwarfStar vs vLLM+DSpark: Running DeepSeek 4 on the RTX Pro 6000
Aug 14, 2026Production benchmarks from our Helix agent fleet on 8× RTX PRO 6000: the engine journey from DwarfStar (122 tok/s) to SGLang (653 tok/s) to vLLM+DSpark (946 tok/s box aggregate, 204 tok/s single-stream). Same hardware, 7.8× throughput.
Read more →What's Actually in the Sovereign Server
Aug 14, 2026We ship a 4U rack server with 8× NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs and 768 GB of VRAM. The same silicon now serves Qwen3.8 at 153.3 tok/s single-stream with 208 scheduler slots, while DeepSeek reaches 1.82–1.89K output tok/s peak decode.
Read more →Automate Your Own Job with AI
Aug 2, 2026Automate your own job with AI - every automation wave since Ford optimised the business, never the individual, so start by reclaiming the first hour of your day.
Read more →The Subtraction Dividend: Helix Now Runs Inference on AWS Trainium
Jun 30, 2026Serve models on AWS Trainium and Inferentia with Helix: cheaper per token, real capacity when GPUs are scarce, and a drop-in OpenAI-compatible endpoint.
Read more →An AI Agent Deleted Our CI System While I Was on the Runway at Heathrow
Jun 10, 2026I gave an AI agent kubectl access to production. It deleted our CI. Now agents open PRs instead — and the first one just fixed a volume alert at 4am.
Read more →The Second-Order Effects of AI-Augmented Organisations
Jun 8, 2026Once AI colleagues are real, span of control collapses, hierarchies flatten, team-size doctrines die, and the new failure mode is attentional drift, not cognitive overload.
Read more →Giving Every Agent Its Own Desktop
Jun 3, 2026I recorded my talk at Tessl AI DevCon on my laptop. Here's what I said about why agents need their own computers, why Claude Code made me stupider, and why multi-agent org charts devolve into corporate politics.
Read more →Modelling the AI-Augmented Organisation
May 26, 2026Swarms of agents is the wrong mental model. Reuse the org-chart and model AI as colleagues with roles, responsibilities, and a human accountable at the top.
Read more →Four Lessons from Building an Agentic Workforce
May 15, 2026Notes from wiring up agents with personalities to streams of information: cascading self-activations, models that won't shut up, emergent enterprise politics, and agents that forget they aren't human.
Read more →Working with the Garage Door Up, Without a Door
May 7, 2026Why every design doc we write goes straight to a public Git branch, and what we've learned by leaving the door off.
Read more →Chunking Beats Slicing: What SWE-Bench Taught Me About Code Intelligence
Mar 17, 2026I benchmarked two approaches to code indexing. The smarter one lost.
Read more →76ms writes on an NVMe: how ZFS dedup ate my agent infrastructure alive
Mar 16, 2026ZFS dedup was causing 4-6x write amplification and 76ms latency spikes on our agent infrastructure. The fix was a ZFS feature from 2005.
Read more →Adding support for Gemini thought signatures
Mar 7, 2026Gemini 3 and 2.5 models require thought signatures for multi-turn tool calling. Here's what they are, why they break OpenAI-compatible proxies, and how we fixed it in Helix with a global signature cache.
Read more →The $175K Server That Replaces Your Cloud AI Bill
Mar 5, 2026We're shipping a 4U rack server with 8× NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, 768 GB VRAM, and Helix preloaded. The maths works — and sovereignty matters more than cost.
Read more →Trump Can Read Your Email — and Your AI Agent Logs: The Case for Digital Sovereignty
Mar 5, 2026I was on a video call with Dubai while missiles were being intercepted overhead. The next day I watched Trump threaten the UK prime minister on live television. It connected some dots about AI infrastructure.
Read more →Introducing the Helix Blog
Feb 25, 2026Welcome to the Helix blog — your source for insights, tutorials, and updates on AI agent orchestration with HelixML.
Read more →