Sage A1.5 in Hermes Agent: a 35B open model run locally with llama.cpp, tested on a Jupiter flyby and parallel sub-agents

The setup used

• Serving: locally with llama.cpp. VRAM use is about 38 GB; the speaker says lowering the context window reduces it.
• Harness: Hermes Agent, which spawns and displays parallel sub-agents. He says Hermes' sub-agent orchestration has improved a lot lately.
• Model: 35B mixture-of-experts, built on Qwen 3.6 and the earlier Sage A1, with four domain skills (search, software engineering, ML engineering, science) and two execution layers on top: orchestration (delegate to parallel sub-agents and integrate results) and verification (gather evidence, review its own judgment and adjust).

Reproduce the test

1. Serve the model locally with llama.cpp. Expect about 38 GB of VRAM and lower the context window if you are short.
2. Use Hermes Agent as the harness, pointed at that local server. The video does not show the Hermes configuration, so follow the Hermes docs for adding a custom endpoint.
3. Give a physics task that forces verification: simulate a spacecraft gravity assist, compare the result with the analytic prediction, and produce an animated HTML replay.
4. Give a delegation task: ask for three named sub-agents to run in parallel and one page assembled from their answers, then watch the sub-agents in Hermes.
5. Check the numbers yourself rather than trusting the report: here the simulated gain and analytic gain were both 9.996 km/s.

Test 1: Jupiter gravity assist

The prompt asked for a Voyager-style Jupiter flyby simulation, a check of the speed gain against the analytic prediction, and an animated HTML replay. After more than an hour and four tool calls using sub-agents it produced a Python script. The simulated speed gain was 9.996 km/s, matching the analytic prediction of 9.996 km/s, and an HTML file replaying the flyby in Jupiter's frame beside the spacecraft speed relative to the Sun. The point of the test is the verification behaviour the model card highlights: simulate, compare to theory, fix.

Test 2: parallel sub-agents

The prompt asked the model to fact-check astrology by delegating to three sub-agents in parallel: an astronomer (where the planets really were), a physicist (does Mars pull on a newborn more than the midwife does) and a statistician (why generic predictions feel true), then assemble one web page. Hermes showed the three running in parallel, one finishing first and another agent verifying. The result: Mars pulls on a newborn less than the midwife does, and about 47% of people find two or more vague predictions that fit by luck. The merged page was titled "fortune telling with science".

How it was trained (model card, as relayed)

• Stage one: separate teacher models for long-horizon search, engineering and scientific research, each with its own training environments; the best trajectories go through supervised fine-tuning then reinforcement learning.
• Stage two: the teachers are merged into one model with a joint post-training step that combines outcome rewards with on-policy distillation, where teachers give token-level feedback on what the student generates.

Caveats

• The model card compares favourably with Kimi K3, GLM 5.3 and a GPT 5.6 model. The speaker says he does not trust benchmarks and these are not independently reproduced here.
• The Jupiter run took over an hour; this is a slow, heavy model for hard science tasks, not a chat model.
• Two tests, one run each, judged by the speaker (and a judge model, whose verdict was not captured here).
• No Hermes configuration was shown in the transcript, so this page does not give Hermes commands. See our Hermes guides, which are checked against the Hermes docs.