ENPT

basal (2.0 + omni)

Last updated:

Open102–102ms$0/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity6.1/10

    Apache-2.0 weights from 0.8B-class to 31B with FP8, GGUF and MLX ports; an engine repo (rkinas/basal, 152 stars) with a /v1/systemone server, Python client, MCP tool and per-model calibration files; a docs site and a technical report. One author, no hosted API, no SLA.

  • Capability8.1/10

    Official JevBench (17 Oct): Intelligence 74.1, calibration 91.0, p50 0.102 s. Choice, Noul, Score plus multi, act and tool.

  • Adoption14/100

    152 GitHub stars, about 7k first-party HF downloads across the 1.0 / 1.5 / 2.0 lines (the 2.0 repos 290-357 each in five days), 30-odd likes, GGUF / MLX / FP8 ports, a third-party runtime (Deqio) listing. A site page for users, no named production case verified.

Vendor claims

publicado (OK do Fabricio 17/10 BRT). Board audit of 17 Oct 2026: see the “Board audit” section below; where it conflicts with older board statements on this page, it wins.

publicado (OK do Fabricio 17/10 BRT).

Read this first.

  1. Only one model has independent evidence. The independent number is basal-omni-1.0-ultra (31B, Gemma-4-31B-it): Benchmark Heaven added it to JevBench on 17 Oct (release v1.6.28) and it leads the open-weights Capability ranking of Jev-class systems at 82.5. The basal-2.0 models (1.5B / 4.5B / 11B) are not on JevBench, the Decision Index 0.3 board or Artificial Analysis (we found no row).
  2. Werdykt is the project’s own benchmark. Its site says so (“autorski benchmark projektu basal”) and that the compared models were pre-selected from the top of the JevBench ranking. Treat every Werdykt number as vendor-run.
  3. basal-omni is experimental. It is trained on text decisions only; images and audio (audio only on the smaller omni sizes) are read zero-shot.

basal (from the Polish basal ganglia) is a family of typed-decision models and an inference engine by Remek Kinas. A model reads a state and returns a calibrated probability for each allowed answer in one forward pass, with no text. It is aimed at Polish and English, and the classes are described in the request instead of being trained in. We catalogue the family as one entry; it was on the radar as a watch item for basal 1.0 / 1.5 since 3 Oct 2026.

The family

Line Models Base Notes
basal-2.0 (12 Oct) mini 1.5B (8,192-token context), 4.5B, max 11B Bielik (Polish) v3 -FP8, -GGUF, -MLX-8bit ports per size
basal-omni-1.0 (10-17 Oct) mini 4B, max 12B, 1.1-max 12B, ultra 31B Gemma 4 (E4B / 12B / 31B) experimental; ultra up to 255 options, text and images; mini/max add audio
Engine rkinas/basal (v2.0.0 for the cards, README says 2.1) Apache-2.0, 152 stars; basal-serve exposes POST /v1/systemone

Question types: choice, noul, score, and in the engine multi, act (with costs and a max_error that defers uncertain cases) and tool (picks among OpenAI/MCP tool definitions). Languages: Polish, English.

JevBench (independent): basal-omni-1.0-ultra

Benchmark Heaven JevBench open-weights board (v1.7.46, 17 Oct 2026; the model arrived in release v1.6.28, same day):

Metric basal-omni-1.0-ultra
Capability Score 82.5 (#1 among Jev-class systems; Quyet-1.0-Large 81.7, Jeff-1.0-Large 81.5)
Intelligence 74.1
Calibration 91.0
Cost $0.045 per 1,000 decisions (1.38× the Jev v1.5 reference)
Median latency 0.102 s measured on-GPU (the board’s estimate incl. network: 0.35 s)

The release notes say it ranked 4 of 45 on the composite and 4 on Capability within that draw. The board lists the base as undisclosed on the model line while the HF card says Gemma-4-31B-it; we use the card. The board’s “Jev-class” cap is at most 2× Jev’s cost and median latency.

Vendor-run numbers

  • Werdykt v1 (5,000 hidden items, own board, 102 systems listed): basal-2.0-max 0.784, ranked 31 of 94 when the card was written; omni-1.0-max 0.823; omni-1.0-ultra 0.879. Own calibration: ECE 0.044 (2.0-max); at a 5% target error 65.6% of items automated at 4.51% observed error.
  • Decision Index 0.2.1: basal-2.0-max 44.07 (own run, 0 of 150,317 unanswered). There is no 0.3 board entry for any basal model.
  • Speed: c1 p50 37 ms, 16.2 decisions/s at 16 clients on one H100 (basal-2.0-max, own pass).
  • The card’s own limits: basal-2.0 is below basal-1.5 on a 10-task English public panel (−1.8 to −3.6 points) and automates fewer decisions at the 1% target; max_error: 0.01 defers nearly everything with the shipped calibration files; act, tool and multi are experimental.

Fit / anti-fit

Fit: Polish-first (and English) routing, policy checks and triage with a confidence threshold that defers to a person, on a single GPU, with Apache-2.0 weights and an auditable calibration file.

Anti-fit: a hosted API or SLA, multilingual needs beyond Polish and English, or relying on Werdykt as proof (it is the project’s own).

Limits

  • One author; the engine README and the model cards sometimes carry different version labels (2.0.0 vs 2.1).
  • basal-omni is experimental and served in eager mode (the ultra FP8 port is text only).
  • The JevBench row exists only for the ultra model; the smaller models have no independent result.

Why it is in the catalog

Public Apache-2.0 weights, a documented typed API in the System One shape with per-model calibration, an active engine repo, and now an independent top Capability result on the JevBench open-weights board.

Score working (17 Oct 2026)

  • Maturity = availability 8 × 0.30 + docs 9 × 0.25 + integrations 4 × 0.25 + support 2 × 0.20 = 2.40 + 2.25 + 1.00 + 0.40 = 6.05 → 6.1
  • Capability = types 10 × 0.30 + calibration 7 × 0.25 + latency 7 × 0.25 + accuracy 8 × 0.20 = 3.00 + 1.75 + 1.75 + 1.60 = 8.10 → 8.1
  • Adoption = engagement 25 × 0.35 + ecosystem 12 × 0.35 + production 2 × 0.30 = 8.75 + 4.20 + 0.60 = 13.55 → 14

Availability 8: Apache-2.0 weights with FP8 / GGUF / MLX ports and an engine repo; no hosted API. Docs 9: README with API reference, install, hardware and format docs, a technical report with a DOI, and honest limits on every card. Integrations 4: /v1/systemone, Ollama / llama.cpp / MLX ports, vLLM mode, an MCP tool and a Python client. Support 2: one author with a project site.

Types 10: choice, noul, score plus multi, act and tool. Calibration 7: shipped calibration files, a stated method and ECE, board Calibration 91.0, but the ECE is the project’s own (precedent: H2O-Lightning, Quyet). Latency 7: p50 0.102 s measured by the board for the ultra, no p99. Accuracy 8: independent JevBench result for the ultra model; the 2.0 line has none, so the entry is scored on its best verified member.

Engagement 25: 152 stars, about 7k first-party downloads across lines, about 30 likes. Ecosystem 12: ports in four formats and a third-party runtime that supports basal. Production 2: a “who uses basal” page exists, but no named case was verified.

Board audit (17 Oct 2026)

publicado (OK do Fabricio 17/10 BRT). Checked against the official boards on 17 Oct 2026: Decision Index 0.3.2 (data of 16 Oct, 19:00 UTC), JevBench and ImageJevBench live. The boards are separate sources and are never averaged. Older board statements and rankings further down this page are superseded where they conflict with this section.

Board Row Result
JevBench open weights (live 17 Oct) basal-omni-1.0-ultra Capability 82.5; Intelligence 74.1; calibration 91.0; p50 0.102 s; $0.045/1k
ImageJevBench v0.3.0 basal-omni-1.0-ultra Capability 60.6; Intelligence 51.5; calibration 69.6; p50 0.63 s; $0.085/1k

Board-status statements on this page to recheck against the table (may be superseded):

  • “Decision Index 0.2.1: basal-2.0-max 44.07 (own run, 0 of 150,317 unanswered).”

Score update (audit, 17 Oct 2026). Capability 8.1 → 8.1 (unchanged) = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 10 × 0.30 + 7 × 0.25 + 7 × 0.25 + 8 × 0.20 = 8.10 → 8.1. Types, calibration and latency come from the earlier working (10 / 7 / 7); only accuracy is re-derived from the board (8; was 8): JevBench Intelligence 74.1 ÷ 63.6 × 8 = 9.3, half-up, capped at 8. Maturity and adoption unchanged.

Hugging Face on 17 Oct (Remek/basal-2.0-max): 290 downloads, 5 likes, licence tag apache-2.0, last modified 2026-10-14.