Published
Updated
Opened October 19, JevBench ranks basal-omni-1.0-ultra first of 184 open systems on Capability at 82.5
Opened October 19, 2026, JevBench ranks basal-omni-1.0-ultra first of 184 open systems on Capability at 82.5. Composite among open weights is 71.8, rank 3. The card prints Werdykt v1 at 0.879. This desk did not run the weights.
Remek Kinas posted basal-omni on October 17, 2026, at 14:15 UTC. The text says the model is first among open weights on JevBench and second among open and closed systems together, ahead of OpenAI, Microsoft, and TypeSafe’s Jev. The chart on that post is the image saved for this page. The 14:19 UTC reply links the open board.
Earlier that UTC day, at 06:58, he posted that basal-omni-1.0 ultra and basal-omni-1.1 max were on Hugging Face. The links in that post are the 2.1 blog, rkinas/basal, and the omni collection.
The card, opened October 19, names a 31B checkpoint on google/gemma-4-31B-it, Apache-2.0, LoRA rank 32 merged into the weights. A request sends a state and typed questions. The server is basal-serve, path POST /v1/systemone. Choice, noul, and score match basal-1.5. The engine also names multi, act, and tool. Up to 255 options. Images go in "media" with --media on. The 31B Gemma 4 file has no audio encoder, so audio returns HTTP 422. The basal-1.5 page keeps the Bielik family.
Werdykt v1 on that card is 0.879, interval 0.871 to 0.887, Polish 0.881, English 0.877. basal-omni-1.0-max is 0.823. basal-1.5-max is 0.773. Held-out Polish decisions, 7,081 items, print 0.979 against max 0.954. An English panel of 10 public tasks, 3,842 items, prints 0.855 against max 0.789. Coverage at a 1% error target is 56.8% of decisions accepted automatically. One H200, --mode eager, bf16, one option order: 80 ms p50 on 200 sequential requests, 8.6 decisions a second with 32 clients. About 62 GB of bf16 weights. The FP8 port is text only on --mode vllm. We did not load the files.
The Benchmark Heaven blog, dated 18 October 2026, prints Capability 82.54, Intelligence 74.06, Calibration 91.02, and says the board added the model in v1.6.28. It also prints that H2O-Lightning-4B v1.1 leads Composite A among open models at 72.52, with ultra second at 71.76, and that kettei-fat leads the mixed Capability ranking at 86.31.
Opened October 19, about 07:20 UTC, the JevBench JSON is revision v1.6.29, latest_release v1.6.29, 184 ranked open-weight systems and 38 ranked API offerings.
| System | Capability | Open cap. rank | Composite | Open composite rank |
|---|---|---|---|---|
| basal-omni-1.0-ultra | 82.54 | 1 | 71.76 | 3 |
| Quyet-1.0-Large | 81.69 | 2 | 71.39 | 5 |
| Jeff-1.0-Large | 81.50 | 3 | 68.59 | 12 |
| H2O-Lightning-4B v1.1 | 75.00 | 10 | 72.52 | 1 |
| Jev 1.13.0 (reference) | 77.13 | 71.49 | API rank 7 | |
| kettei-fat (API) | 86.31 | 77.80 | API rank 1 | |
| Jev 1.14.0 (API) | 75.01 | 70.20 | API rank 9 |
The ultra row is tagged measured_on 2026-10-17, measured_in v1.6.28, draw v1.6-regular-20261012, status anchor-runs-pending. Intelligence 74.06, Calibration 91.02, Speed 87.31, Cost 50.51. Raw p50 0.102 s on GPU loopback. Adjusted p50 0.354 s, x2 + 0.15 s. Estimated $0.04465 per 1,000 decisions from a frozen Gemma 4 31B tariff, not a basal bill. Torchcast Decision 4B v1.2 sits at open composite rank 2, 71.84. We did not rerun the suite.
Access does not list basal-serve as a door for calling Jev.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Remek Kinas, @KinasRemek, October 17, 2026, 14:15 UTC, status 2111461103322452227. The text says basal-omni is first among open models on JevBench and second among open and closed together, ahead of OpenAI, Microsoft, and TypeSafe's Jev. The share image on this page is the chart from that post. The 14:19 UTC reply, status 2111461989729829173, links https://benchmarkheaven.com/jev-models. The 06:58 UTC post, status 2111351001177583781, says basal-omni-1.0 ultra and basal-omni-1.1 max are on Hugging Face, and links the 2.1 blog, GitHub, and the omni collection.
The Hugging Face card Remek/basal-omni-1.0-ultra, opened October 19, names a 31B Gemma 4 checkpoint, Apache-2.0, from google/gemma-4-31B-it with a rank-32 LoRA merged into the weights. Werdykt v1 macro is 0.879, 95% interval 0.871 to 0.887, Polish 0.881, English 0.877. The 12B max is 0.823. basal-1.5-max is 0.773. The 31B base has no audio encoder. Images are zero-shot through Gemma 4's encoder. POST /v1/systemone on basal-serve. One H200, --mode eager, bf16, one option order: 80 ms p50 sequential, 8.6 decisions a second with 32 clients. About 62 GB of bf16 weights. We did not load them.
Opened October 19, 2026, about 07:20 UTC, JevBench JSON revision v1.6.29, latest_release v1.6.29, 184 ranked open-weight systems and 38 ranked API offerings. basal-omni-1.0-ultra Capability 82.54, first of 184 open systems on that axis, Intelligence 74.06, Calibration 91.02. Composite 71.76, open composite rank 3. Measured on 2026-10-17, draw v1.6-regular-20261012, measured_in v1.6.28, status anchor-runs-pending. Raw p50 0.102 s on GPU loopback, adjusted 0.354 s (x2 + 0.15 s). Estimated $0.04465 per 1,000 decisions. H2O-Lightning-4B v1.1 is open composite rank 1 at 72.52. Torchcast Decision 4B v1.2 is rank 2 at 71.84. Quyet-1.0-Large is Capability rank 2 at 81.69. Jeff-1.0-Large is rank 3 at 81.50. kettei-fat stays API composite rank 1 at 77.8, Capability 86.31. Jev 1.13.0 is a reference row, composite 71.49, API rank 7, Capability 77.13. Jev 1.14.0 is API rank 9 at composite 70.2, Capability 75.01. We did not rerun the suite.
Compare
82.5 is Florian's open-weights Capability score on revision v1.6.29, 1,500 items, a v1.6.28 measurement put on that export. Remek's October 17 post printed TOP1 among open models. The JSON we opened still ranks it first of 184 on Capability. Composite 71.76 is a different ranking. H2O-Lightning-4B v1.1 at 72.52 is first on that open composite table.
0.879 is Werdykt v1 on the ultra card, 5,000 hidden Polish and English decisions. It is not JevBench 82.5. basal-1.5's October 4 Werdykt cell 72.12% against Jev 1.13.0 at 81.64% stays on that page. The 2.1 blog prints Jev 1.14 at 0.822 on Werdykt, which is a later row on basal's own set.
80 ms p50 is the card's sequential H200 load. JevBench's 0.102 s raw p50 is the board's GPU loopback. The adjusted 0.354 s applies x2 + 0.15 s. Do not merge those three clocks.
Access on this desk does not list basal-serve as a door for calling Jev.
Terms
- basal-omni-1.0-ultra
- Remek Kinas's Apache-2.0 31B checkpoint on Gemma 4, served by basal-serve at POST /v1/systemone. Opened October 19, 2026, JevBench ranks it first of 184 open systems on Capability at 82.5. The card prints Werdykt v1 at 0.879. Text and images. No audio.
- Capability Score
- JevBench's mean of Intelligence and Calibration, ranked among systems inside twice Jev's cost and median latency caps. Opened October 19, those caps are about $0.0646 per 1,000 decisions and 1.23 s.
- Werdykt v1
- basal's hidden 5,000-decision set, 10 categories in Polish and in English. The ultra card prints 0.879. The questions are not published.