SIQ-1 Qwen3.6 Achieves State‑of‑the‑Art Autoresearch Performance via PPO
A new open‑weight model has pushed the frontier of local AI research: SIQ-1, a PPO‑fine‑tuned version of Qwen‑35B‑A3, demonstrates verifiable gains on autonomous‑research benchmarks that rival or exceed those of much larger proprietary sys…
A new open‑weight model has pushed the frontier of local AI research: SIQ-1, a PPO‑fine‑tuned version of Qwen‑35B‑A3, demonstrates verifiable gains on autonomous‑research benchmarks that rival or exceed those of much larger proprietary systems [3]. This result is notable not only for its raw performance but also for what it signals about the viability of independent, self‑hosted agents that can conduct meaningful research without relying on massive cloud‑scale infrastructure.
SIQ-1 begins with the Qwen‑35B‑A3 base, a 35‑parameter‑billion model already known for strong reasoning and code abilities. The author applied proximal policy optimization (PPO) with a carefully shaped reward function, noting that this is the first time they have observed PPO “actually pull its weight” with a measurable signal [3]. The training loop therefore produced a model that not only improves on next‑token prediction but learns to generate higher‑quality research ideas when prompted to autonomously explore a problem space.
On the karpathy/autoresearch benchmark for parameter‑golf — a task that measures how compactly a model can propose effective architectural changes — SIQ-1 outperforms both GLM-5.2 and the considerably larger Qwen‑350B [3]. The qualitative impression is that the generated suggestions “feel Opus4.8‑like,” indicating a level of creativity and technical depth typically associated with the most advanced closed models. Separately, on the bullshit‑bench — a probe designed to test resistance to vacuous or hallucinated outputs — SIQ-1 bests NEX and GPT‑5.5 [3]. These two complementary evaluations show that the model advances both productive idea generation and factual discipline.
The model and its associated agent demo are publicly available: the GGUF‑quantized weights sit on Hugging Face under AlexWortega/SIQ-1-35B, and a ZeroGPU‑hosted interactive space lets users steer the HERMES agent in real time [3]. This openness means anyone can download, run, and further adapt the model on modest hardware, reinforcing a growing trend where local models are no longer toys but serious research partners.
Why does this matter today? The result dovetails with several parallel developments that collectively shift the balance toward independent AI workflows. First, GLM-5.2 — highlighted for its explicit focus on long‑horizon tasks — has proven itself a strong open alternative, and its max variant is presently ranked