Bobcat 1.3

Last updated: October 17, 2026

Open64–767ms$0/M input

Quadrant scores

See the full quadrant

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

Maturity5.4/10
Apache-2.0 weights (~62 GB BF16, plus an FP8 repo), a reproducible quickstart (vLLM server, official typesafe-sdk works by changing the base URL), GitHub code and a long card. One author, no hosted API, no SLA, no third-party integration.
Capability8.1/10
Choice (1-255), Noul, Score; Full 62.83 on the official Decision Index 0.3.2 (#10; Jev 60.11); median 64 ms / p95 767 ms measured by the maintainers on one RTX PRO 6000. Temperatures per type and ECE 0.015 are the author's own, on a Korean product set.
Adoption2/100
Two days of traffic: 73 + 39 downloads on the weights and FP8 repos, 0 likes, one community GGUF with 732 downloads, 0 GitHub stars. No launch post found, no integrations, no reported users.

Vendor claims

Sealed final on four Korean product tasks (1,609 decisions, opened once): 94.49% accuracy, log-loss 0.335, ECE 0.015; wrong answer planted in the state wins 6.4% of 900 attack questions[Vendor claim — not independently verified]
Decision Index 0.3 public suite, half of its items, own run of the official kit: 59.18; JevBench public items (231), own reconstruction of the v1.6 method, unofficial: 77.2[Vendor claim — not independently verified]
One decision (512 tokens, 8 candidates) p50 28.6 ms on one B200 with vLLM 0.30.0, engine only, no HTTP[Vendor claim — not independently verified]

publicado (OK do Fabricio 17/10 BRT).

Read this first. Bobcat’s headline is an independent number: the maintainers adopted its public run and ran the private tests, giving Full 62.83, #10 (Jev 1.13: 60.11). Most other figures on the card (94.49% sealed final, 92.95% development, the 77.2 JevBench reconstruction) are the author’s own and use Korean product tasks, so they stay vendor claims. The card says plainly that “Jev figures are never used; Jev was never called”.

Bobcat 1.3 is a typed-decision model from an independent author (sanghwa-na, code at foxl-ai/bobcat): Gemma-4-31B-it with three rank-16 LoRA arms averaged in weight space. You send a state and named candidates; it returns the top name and a probability for every name, with no generated text. Languages: English and Korean.

Specs

Attribute	Value
Developer	`sanghwa-na` (independent); GitHub `foxl-ai/bobcat` (Apache-2.0)
Base	Gemma-4-31B-it
Weights	Apache-2.0, ~62 GB BF16; FP8 checkpoint (~31 GB on GPU); community GGUF
Types	Choice (1 to 255 named candidates), Noul (P(true)), Score (1 to 10 ordered levels)
Interface	Shapes of TypeSafe’s System One HTTP API; the official `typesafe-sdk` works by changing the base URL; inputs over the limits get HTTP 422, never truncated
Serving	vLLM 0.30.0 plus `bobcat.flash_server`; one RTX PRO 6000 96 GB, or one H100 80 GB with an FP8 KV cache
Context	98,368 tokens in the quickstart (states up to ~96K); the card reports 0 failures on 240 questions up to 96,161 tokens
Calibration	One temperature per question type (choice 1.0078, noul 0.4454, score 1.0234); a single 1.2227 alternative
Languages	English, Korean
Released	11 Oct 2026 (card updated 14 Oct)

Official board: Decision Index 0.3.2

Data generated 16 Oct 2026 19:00 UTC. Full = 0.2 public + 0.5 same-skills + 0.3 new-domains.

	Bobcat 1.3	Jev 1.13
Full	62.83 (#10)	60.11 (#18)
Public	59.31 (#27)	57.96 (#37)
Requests answered	110,201 of 110,201
Single-request latency, median / p95	64 / 767 ms (maintainers, one RTX PRO 6000, the quickstart’s FP8 serving)

The maintainers adopted Bobcat’s public 0.3 run (their rescore of 59.31 equals the author’s claim) and ran the private tests. Not on JevBench or Artificial Analysis as of 17 Oct (we found no row).

What is vendor-run

Sealed final 94.49%, development 92.95%, TypeSafe workflow examples 92.10%: all on the author’s own tasks; the product evaluation is Korean and four-task.
Decision Index public (half): 59.18 and JevBench public items 77.2: the author’s own runs; the JevBench figure rebuilds the v1.6 method and is unofficial. Do not mix it with Benchmark Heaven’s board.
ECE 0.015 (sealed final) is the author’s measurement with the shipped temperatures.
p50 28.6 ms is an engine-only B200 figure with no HTTP; the board’s 64 ms median is the number to use.

Fit / anti-fit

Fit: a self-hosted, Apache-2.0 decision model with an independent top-10 result, a System One-shaped server, Korean and English, and long (96K) states.

Anti-fit: a hosted API, support, a track record (two days old, one author), or very low latency tails (p95 767 ms).

Limits

One author; the repo was pushed on 14 Oct and has 0 stars.
Calibration temperatures were fitted on Bobcat’s own held-out splits and on LocalLLaMA typed-decisions rows; the card says a temperature never changes an argmax.
The card reports a 6.4% rate of a planted wrong answer winning (300 questions × 3 attacks).

Why it is in the catalog

Public Apache-2.0 weights, a typed Choice / Noul / Score interface in the System One shape, and an independent official-board result above Jev’s Full score. It passes the inclusion bar.

Score working (17 Oct 2026)

Maturity = availability 8 × 0.30 + docs 8 × 0.25 + integrations 3 × 0.25 + support 1 × 0.20 = 2.40 + 2.00 + 0.75 + 0.20 = 5.35 → 5.4
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 3 × 0.35 + ecosystem 3 × 0.35 + production 0 × 0.30 = 1.05 + 1.05 + 0.00 = 2.10 → 2

Availability 8: Apache-2.0 weights with an FP8 repo and a community GGUF, but custom serving code. Docs 8: a long, reproducible card with manifests and per-task tables. Integrations 3: /v1/systemone-shaped server (precedent). Support 1: one author.

Calibration 7: a published method and ECE, vendor-run (same rule as pplx-decider v1.2). Latency 7: median 64 ms measured by the maintainers, p95 767 ms. Accuracy 8: independent board result (Full 62.83, #10).

Engagement 3 and ecosystem 3: 112 downloads on the main repos, 0 likes, one community GGUF (732), no post. Production 0.