Published
Updated
Rizzo Flow is 49 ms on an RTX 5060 Ti, and six missing-evidence rows stay confident
This topic was created 31 days ago, and the information it contains may have evolved or changed since then.
Simone Rizzo's Apache-2.0 server reads Spark-X2.5 logits through llama.cpp and speaks POST /v1/systemone. On Windows with an RTX 5060 Ti, the 4B Q8_0 checkpoint is 49 ms at the median and uses 5.6 GiB. On SemIf's authored144 fixture the balanced accuracy is 0.812 against SemIf's published 0.819. Where the right answer is insufficient evidence, 6 of 36 rows are still confident and wrong.
Geek Lite posted on September 23 that Rizzo Flow puts an open model behind the same HTTP shape as Jev. The post’s figures match the README: probabilities are uncalibrated by default, 6 of 36 answers stay confident when the evidence is missing, each question has at most 26 options against Jev’s 255, and 49 ms is a Windows run on an RTX 5060 Ti. The author on the README is Simone Rizzo, of Rizzo AI Academy. The license is Apache-2.0. The project says it is not affiliated with TypeSafe and does not reproduce Jev’s training.
The default weights are Spark-X2.5-4B at Q8_0, about 4.4 GB, through llama.cpp release b11081. A 1.7B checkpoint is optional. rizzo serve listens on port 8017 and accepts POST /v1/systemone. A jev-* model name is accepted as an alias. The response always reports the local model id. Output tokens are 0. The state is prefilled once. Each question reads the logits of the answer letters and stops. Confidence is (n * p_max - 1) / (n - 1). The README says that is the statistic on TypeSafe’s confidence page, and that Jev’s exact formula is not public.
The current table is Windows 10 with an RTX 5060 Ti, 16 GB, CUDA, prompt v3. Short decisions at Q8_0: p50 49 ms, p95 52 ms. Peak GPU memory 5.6 GiB. A shared-prefix run of shape777 (37 states of about 2,000 tokens, 21 criteria each) is 20.99 decisions a second, 1.00 seconds per state. Scoring the same questions fresh is 2.60 decisions a second. Between those two modes, 13 of 777 answers change, and the largest probability gap is 0.163. The README says every one of those flips had a margin under 0.24, and that a threshold should not sit near 0.5. Vulkan on the same card is 90 / 94 ms and agrees with CUDA on all but 3 of 252 rows. Linux, Mac, AMD, Intel, and CPU-only are listed as not tried for this runtime.
Quality is scored on SemIf’s fixtures with SemIf’s evaluator. authored144, mean-family balanced accuracy: Q8_0 0.812, held-out half of 72 rows 0.793. SemIf’s published Q8 cells are 0.819 and 0.811. The paired difference versus SemIf is -0.007, interval -0.076 to +0.065. perturbations108 is 0.848 against SemIf’s 0.766. On 36 rows where the right answer is that the evidence is insufficient, accuracy is 0.750, and 6 answers at probability 0.8 or higher are confidently wrong. SemIf’s published count of that miss is 1. The 1.7B checkpoint is 0.678 on authored144 and 25 / 27 ms a decision. The README says half of authored144 is the dev split the prompt was chosen on.
A Snake page records 140 moves in 25.6 seconds, about 150 ms a decision, on an earlier MLX runtime. Three informal games are described. The README calls them a handful of games, not a benchmark. Sixty-five tests run without weights. We did not start the server.
SemIf is the open logit readout this table is tied to, on Qwen3.5-4B. Rizzo Flow swaps in Spark-X2.5 and llama.cpp, and the authored144 gap against SemIf’s published row includes zero. The missing-evidence miss does not: 6 confident wrong answers against SemIf’s 1.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Primary post is @QingQ77, September 23, 2026, 02:05 UTC. Display name Geek Lite. The post summarizes the repository: open weights plus llama.cpp, a localhost base URL, uncalibrated probabilities by default, 6 of 36 confident wrong answers when evidence is missing, 26 options against Jev's 255, and 49 ms on the author's Windows machine with an RTX 5060 Ti.
README author is Simone Rizzo, Rizzo AI Academy, Apache-2.0. Default model Spark-X2.5-4B Q8_0, also 1.7B. llama.cpp release b11081. Current numbers: Windows 10, RTX 5060 Ti 16 GB, CUDA.
authored144 mean-family balanced accuracy 0.812 at Q8_0, held-out half 0.793. SemIf's published Q8 cells on the same rows: 0.819 and 0.811. perturbations108: 0.848 against SemIf 0.766.
Missing evidence, 36 rows: accuracy 0.750. Confident answers at p at least 0.8 where insufficient was right: 6, against SemIf's 1. Short-state latency p50 / p95: 49 / 52 ms. shape777 shared: 20.99 decisions per second, 1.00 s per state.
Shared versus fresh argmax changes: 13 of 777, max probability gap 0.163. Peak GPU memory 5.6 GiB. The 1.7B Q8_0 row is authored144 0.678, 25 / 27 ms. Paired difference versus SemIf on authored144 at Q8_0: -0.007, interval -0.076 to +0.065.
At most 26 answer slots. Jev's documented cap is 255. SemIf's is 16. Mac, Linux, AMD, Intel, and CPU-only are untested on this runtime. Probabilities are uncalibrated unless the user fits a calibration.
65 tests without weights. Independent of TypeSafe. We did not run the server.
Compare
SemIf is the fixture this README scores against, and the paired gap on authored144 includes zero. rev and spark-s1 score public JevBench tiers instead. Those headers are 48, 72, and 111.
Rizzo Flow does not print a JevBench composite. The 6-of-36 miss is the same missing-evidence file SemIf published, where SemIf's count is 1.
Terms
- answer slot
- Rizzo Flow maps each candidate to one uppercase letter that must be a single token. Twenty-six letters means at most 26 options. Abstention uses a slot too. The README contrasts that with Jev at 255 and SemIf at 16.
- authored144
- SemIf's committed fixture, scored with SemIf's own evaluator. Rizzo Flow Q8_0 on CUDA is 0.812 balanced accuracy, 0.793 on the held-out half. SemIf's published Q8 cells are 0.819 and 0.811. Half of the rows are the dev split the prompt was chosen on.