Published
Updated
On Ámbar Reflex the 4B logit player went further than the more accurate 9B
This topic was created 31 days ago, and the information it contains may have evolved or changed since then.
Abel Coronado's local server plays an original run-and-gun with six typed questions and no generated text. Qwen3.5-4B at 8-bit on an RTX 3060 advances 24.9% of the level in 408 ms a decision. The same size model writing JSON advances 10.3% in 1,041 ms. Offline, Qwen3.5-9B is right on 82.7% of 554 labelled decisions and reaches 17.5% when the clock is running.
Abel Coronado posted twice on September 23 about a local server that plays Ámbar Reflex, an original 8-bit run-and-gun. The first post: Qwen3.5-4B at 8 bits on an RTX 3060 decides in 408 ms and reaches 24.9% of the level, without generating text. The same size model writing JSON takes 1,041 ms and reaches 10.3%. The second post puts three local models on the same match. Qwen3.5-9B is the most accurate at 82.7%. Qwen3.5-4B is the one that gets further, at 24.9%. The repository is abxda/ambar-reflex-sistema-1, MIT. The README says the game was built with Claude Opus 5.5 and GPT-6 Astra under Coronado’s direction, and that the project is not affiliated with TypeSafe.
Each decision is one POST /v1/systemone with six questions: move, jump, duck, aim, shoot, and danger. The server reads logits. The README credits that method to SemIf, on Qwen3.5-4B Q8_0 for the shipped player. The TypeSafe SDK is described as working unchanged with TYPESAFE_BASE_URL pointed at 127.0.0.1:8765.
The realtime table in the README: typed decisions, N=5, mean progress 24.9%, 408 ms a decision. The same size model writing JSON, N=3, 10.3% and 1,041 ms. A random agent, N=5, 4.6%. Offline, with no time pressure, 554 labelled decisions: Qwen3.5-4B Q8_0 is 77.8%, Gemma 4 E4B at Q6_K is 81.6%, Qwen3.5-9B at Q5_K_M is 82.7%. On the clock those three reach 24.9% at 408 ms, 14.4% at 312 ms, and 17.5% at 591 ms. Gemma jumps on 74% of decisions. The README says the questions were tuned with the 4B, so that model is playing at home.
The benchmark notes a GPU freeze, NVRM Xid 56, when the same card was also drawing the desktop. A watchdog aborts on Xid errors, above 80 °C, or low free VRAM, and the instructions say to run one match at a time. We did not run it.
jev-rubiks asked hosted Jev which cube turn to take, and the distance stayed at 22. This table is local weights under a frame clock. The accuracy winner and the progress winner are different rows.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
First post is @abxda, September 23, 2026, 01:10 UTC. Display name Abel Coronado, Ph.D. Text: local Jev clone, Qwen3.5-4B at 8 bits on an RTX 3060, no generated text, 408 ms and 24.9% of the level in a run-and-gun; writing JSON, 1,041 ms and 10.3%.
Built with Opus 5.5 and GPT-6 Astra. Second post, 02:37 UTC: three local models, no thinking, same match. Qwen3.5-9B is the most accurate at 82.7%, and Qwen3.5-4B goes further at 24.9%.
README: MIT. Sistema 1, typed decisions, realtime, N=5: mean progress 24.9%, 408 ms. Sistema 2, same size writing JSON, realtime, N=3: 10.3%, 1,041 ms. Random agent, N=5: 4.6%. Offline accuracy on 554 labelled decisions, no time pressure: Qwen3.5-4B Q8_0 77.8%, Gemma 4 E4B Q6_K 81.6%, Qwen3.5-9B Q5_K_M 82.7%.
Realtime progress and latency for those three: 24.9% at 408 ms, 14.4% at 312 ms, 17.5% at 591 ms. Gemma jumps on 74% of decisions. The README says the questions were tuned with the 4B, which plays at home.
Six questions per request: move, jump, duck, aim, shoot, danger. Logit method credited to SemIf. Server is POST /v1/systemone. They report an NVRM Xid 56 freeze when the GPU also drew the desktop.
We did not run the game.
Compare
jev-rubiks asked hosted Jev to pick a cube turn and the distance stayed at 22. Tetris demos hand Jev a legal-move list. This loop is local weights, a clock, and a progress percent.
The 9B model wins the 554-decision accuracy table and loses the realtime progress table to the 4B.
Terms
- Sistema 1
- Coronado's name for the typed path. Six closed questions in one POST /v1/systemone, probabilities only, no generated text. On the realtime table, N=5, mean progress is 24.9% at 408 ms a decision.
- 554 decisions
- The offline accuracy set, no time pressure. Qwen3.5-4B Q8_0 is 77.8%, Gemma 4 E4B Q6_K is 81.6%, Qwen3.5-9B Q5_K_M is 82.7%. The README says the questions were tuned on the 4B.