The Pulse, Monday 5 October 2026
Aleph Alpha released Kolibri, a sovereign open-weight German LLM with real traction on HN. Claude Code added the "You Should Know" plugin to catch important output you might miss. DuplexJev proposes batched speech decisions without decoding — 80 decisions per second on an 8-GPU node, directly useful for full-duplex voice agents.
Models
inclusionAI: Ling 3.1 Flash
A hybrid reasoning mixture-of-experts model with 25B active parameters out of 560B total, now on OpenRouter.
Worth knowing, no direct use for us yet.
Google DiarizationLM-Gemma-4-E4B-v1
A Gemma-4 based speaker diarization model released on Hugging Face.
EverRaise's voice agents on Vapi need accurate speaker identification in multi-party calls.
Research
Batched Speech Decisions Without Decoding: Single-Token Supervision Lets a Frozen LLM Hear Beyond the Transcript
DuplexJev feeds ASR-encoder hidden states through a small connector into a frozen LLM and reads each question as a single-token distribution over options, answering 80 decisions per second on an 8-GPU node without any decoding.
EverRaise's full-duplex voice agents make many small real-time decisions; this approach could drastically cut latency.
OpenGameEval: Benchmarking Agentic Programming and Exploration in a Stateful Game Engine
A benchmark that turns Roblox Studio into an executable planning environment with reproducible, stateful game-engine sessions and executable checks.
Relevant benchmark structure for our own agentic code evaluation work and wehshi review panels.
Trained Agentic Context Management
Fine-tunes Qwen3.6-35B-A3B with tools to call itself with any prompt and read tokens from input context in a range, treating context management as an agentic skill rather than a native model capability.
Hermes's context management for long tasks could benefit from this approach rather than packing everything into one window.
APDMem: Agent-Controlled Progressive Disclosure for Query-Adaptive Long-Term Memory
A hierarchical long-term memory architecture that applies progressive disclosure to memory retrieval, adapting to query complexity rather than using flat Top-K.
Botterfly's Mnemosyne memory system could benefit from hierarchical retrieval rather than always returning fixed-count results.
Releases
Python 3.15.0rc3
Release candidate 3 for Python 3.15.0, published October 2.
Every Python interpreter release matters for our services; we test release candidates before production upgrades.
langchain-text-splitters 1.1.3
Dependency updates including tornado, urllib3, anyio, soupsieve, and lxml.
EverRaise and Mackinac both use LangChain for retrieval pipelines; dependency updates sometimes fix memory leaks or security issues.
Langfuse v4.50.0
Adds facet type evaluators, built-in protection, rule idle time, seven-day execution health in evaluator status, and various fixes.
We use Langfuse for observability across SalesKick and Pipeline clients; evaluator health monitoring is directly useful.
Hacker News
Kolibri: A Sovereign Open-Weight Model
Aleph Alpha released Kolibri, an open-weight German LLM with 659 points and 326 comments on HN.
Worth knowing; no direct use yet, but sovereign models show how regional players are competing outside the big three.
Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s
A project demonstrating 125B parameter model inference at 100 tokens per second on a single RTX 4090.
On-premises inference for clients with data residency requirements; shows what's now feasible on consumer hardware.
We're going to need default hard budget caps on pretty much everything
Simon Willison argues for default hard spend limits on API services to prevent runaway costs from agentic systems.
SalesKick's agentic flows and EverRaise's voice agents could rack up unexpected bills; budget caps deserve a place in our deployment checklist.
Getting the most out of Opus 5.5 in Claude and Claude Code
Official guide on using Opus 5.5, with 232 points and 157 comments.
Our team uses Claude Code daily; knowing how to get the best results from Opus 5.5 improves everyone's coding speed.
Agents don't need memory, they need documentation
A post arguing structured documentation beats memory for agent context, with 354 points and 238 comments.
Challenges our investment in Mnemosyne; the counter-argument is worth reading and the HN thread has smart takes on when each approach wins.
People on X
Claude Code's "You Should Know" plugin
ClaudeDevs announced a new plugin that scans Claude's output for important information you might miss, enabled with /plugin enable cc-plugin-you-should-know@builtin.
We use Claude Code on every project; this plugin helps catch decisions and warnings we'd otherwise scroll past.
Linus Tech Tips discusses Hermes Agent
Nous Research noted that Linus Sebastian discussed his Hermes setup experience on their latest podcast.
Hermes is our agent runtime; seeing it reach mainstream tech creators means support and plugin ecosystems will grow.
Sebastian Raschka: Reasoning from Scratch, Round 6
An introduction and implementation of RLVR and GRPO for reasoning models.
Reasoning traces underpin our code review workflows; GRPO is worth understanding if we ever fine-tune a reasoning model ourselves.