Fine-tuned from Qwen3.8-27B on 127 real engineering sessions, it ties its base at 133/150 on BenchLocal while generating 25% fewer tokens, and its multi-token-prediction draft head roughly doubles code generation on an RTX 4090 (49 to 96 tokens/sec). Apache 2.0 BF16 and GGUF weights are on Hugging Face, and the benchmarks are Continuity Research's own runs.
The MIT repo, built from the setup behind Votito, bundles a sandboxed Claude Code container, a bash command regulator that auto approves or denies with guidance, a lint on edit PostToolUse hook, and a generator that turns a plan into multi agent workflow scripts with test gates. Its author says moving that orchestration out of the LLM cut token spend by about 80% versus a bare Claude workflow (Show HN)…
The guided setup learns from your project and recent sessions and now also runs on Bedrock, Vertex, Foundry, and telemetry disabled sessions, while the new /claude api managed agents secure checks a Managed Agents config against Anthropic's security guide without changing anything. Pro and Max users also get a warning before resuming an idle conversation would eat a noticeable share of their 5 hour usage limit…
The new coding CLI has Agent, Build (parallel tasks gated by tests), and Architect modes; Architect turns an approved blueprint into a plan its Agent OS compiler hands to parallel agents, and Spinoza's demo is a 173,773 line finance terminal built from 1,111 commits. Pro Architect costs $200/month, which per the founder on HN includes $200 in token usage…
Built on the Pi harness, Fade runs a conversational companion and a Builder side by side so you can plan or line up tasks mid build, and it only marks work "checked" after committing the change and re running the Builder's own test or command proofs itself. It's an early, self described rough v0.5 (MIT, Node 22.19+, works with any Pi provider including Ollama), commits straight to your current branch, and has a no key npx @muzaffermahi/fade demo to try first…
The RL trained model treats a trace like a repo, grepping spans to catch quiet errors such as an agent marking work done while tests still fail, and on Laminar's own 523 trace benchmark it edges GPT 6 Sol on detection F1 (.835 vs .816) at 23x lower cost, though Claude Opus 5 still scores higher (.890). Laminar's observability platform is Apache 2.0 and self hostable, and flow 1 runs in its Signals agent at $0.05/$0.30 per million input/output tokens (benchmark)…
Spotted something we missed? Start a thread.