What a difference a year makes #
We just held meetup #61 on 2026-10-27 in Halle (Saale). About a year ago, we had a drinkup #54 on 2025-09-30. The wrap-up blog post for #54 demos the crush agent running qwen3-coder:30b with ollama, on an RTX 4000 SFF “Ada” GPU with 20GB VRAM.
If you squint, you can see a tiny bit of agentic coding happening. Claude code was released about seven months earlier on 2025-02-24, and accelerated the move to agentic coding. Harnesses were built, models improved.
Today, you can run the pi or zot (Go) agent harness with e.g. with the open source strix halo engine gufo (or halogen (LICENSE)) with Qwen3.8-Flash-Next on an AMD Ryzen AI Max 395+ in a Framework Desktop (FWD) with 128GB LPDDR5x-8000:
| Category | Benchmark | Qwen3.8-Flash-Next | Qwen3.8-27B | Qwen3.7-Plus | DeepSeek-V4-Flash-0731 | Claude-Opus-4.6 (Max) |
|---|---|---|---|---|---|---|
| Model | Params | 125B | 27B | 397B | 284B | – |
| Model | Activated params | 6B | 27B | 17B | 13B | – |
| Model | N-gram embedding params | 51B | – | – | – | – |
| Coding | Agentic coding — DeepSWE 1.1 | 58.7 | 42.2 | 16.5 | 54.4 | – |
| Coding | Agentic coding — SWE-bench Pro | 62.5 | 61.7 | 55.8 | 56.0 | 53.4 |
| Coding | Multilingual software engineering — SWE-bench Multilingual | 81.0 | 73.8 | 75.8 | – | 77.5 |
| Coding | Repo-level code generation — NL2Repo-Bench | 48.1 | 42.3 | 41.1 | 54.2 | 47.6 |
| Agent | Long-horizon office work — CoWorkBench | 73.9 | 70.7 | 65.1 | 45.1 | 68.2 |
| Agent | Professional job tasks — JobBench | 55.7 | 33.4 | 27.6 | 41.3 | 36.6 |
| Agent | Frontier agentic tasks — Agents’ Last Exam (Pass@1) | 24.3 | 20.4 | 13.2 | 25.2 | – |
| Agent | Frontier agentic tasks — Agents’ Last Exam (Score) | 51.2 | 42.9 | 33.6 | – | – |
| Agent | Real-world tool use — Toolathlon Verified (Pass@1) | 73.5 | 67.1 | 50.6 | 70.3 | – |
| General | Instruction following — IFBench | 81.3 | 79.5 | 79.1 | 79.2 | 62.5 |
| General | Scientific reasoning — GPQA Diamond | 91.7 | 89.2 | 90.3 | 90.8 | 91.3 |
| General | Multidisciplinary reasoning — HLE | 35.9 | 30.8 | 34.7 | 33.8 | 40.0 |
| General | Competitive coding — LiveCodeBench v6 | 91.9 | 90.3 | 89.6 | 90.6 | 88.8 |
The agent takes a likely underspecified task and in about fifteen minutes in a multi-turn conversation builds a 1000+ SLOC toy inverted index project. The gif/mp4 shows it in real-time. Energy consumption of the FWD under load stays below 200W (estimated 1c in energy costs).
