Ornith 1.5 and Tiel-Coder, two community fine-tunes of Qwen3.6-35B-A3B, just decisively beat the original 35B-A3B in a 300+ hour tool calling test run by a Reddit developer—outperforming the Qwen3.6-27B dense model and approaching the unreleased Qwen3.8-27B full power. The entire test suite ran on hardware that amounts to a single 32GB V100.

What this is

Qwen3.6-35B-A3B is a small MoE model open-sourced by Alibaba's DAMO Academy in the second half of last year. "MoE" (Mixture of Experts) means the total parameter count is large (35B), but only a small subset (around 3B) activates per inference—so hardware requirements are far lower than dense models of comparable size. The three fine-tunes compared are KAT-Coder, Ornith 1.5, and the newer Tiel-Coder (Tiel is built on Ornith). "Tool calling" is the core capability that lets AI Agents invoke external APIs, operate software, and query databases—nearly every Agent product (whether called Copilot, Agent, or anything else) competes on this at the base layer. The test platform is the open-source llama.cpp local inference framework plus the tool-eval-bench scoring tool, running 65 trials under 128k context pressure.

Industry view

The optimistic read: open-source small models, after community fine-tuning, can now approach top-tier dense models on tool calling—a core Agent capability—on a single 32GB GPU. The inflection point where enterprise deployment costs drop from "cluster" to "workstation PC" may have arrived. But the cautionary counter: this is a result from a single test scenario, and the fine-tuning community is highly fragmented with uneven quality—Ornith-1.5-Heretic's failure within the same test batch illustrates the risk; the original author also stressed that the Qwen3.8 series official version will "basically not be released," meaning this fine-tuning path won't get vendor backing in the short term. Our judgment: this is an early signal of practical local Agents, but it's still at least one version of stability testing and engineering tooling short of enterprise production readiness.

Impact on regular people

  • Enterprise IT: Tool calling capabilities that previously required A100/H100 clusters can now run on a single 32GB GPU—private deployment costs may halve, but the supporting monitoring, rollback, and security audit systems haven't caught up.
  • Individual professionals: Not an actionable signal yet unless you're doing AI development; what's worth watching is that model selection authority is shifting from people who "understand tech" to those who "understand business + can tune models."
  • Consumer market: No consumer-facing products in the short term; in the medium term, the technical feasibility of running Agent tool calling on edge devices (phones, PCs) has improved, but actual deployment still depends on specific use cases.