ENPT

gutsy-0.8b

Last updated:

Open—$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 GGUF files (Q8_0 775 MB, Q4_K_M 505 MB) that run on CPU through llama.cpp; README, model card, release notes and a tested runtime with POST /v1/systemone (works with JevBench's typesafe adapter). No safetensors, no PyPI package. One maintainer, no SLA.

  • Capability5.7/10

    Official JevBench (17 Oct): Intelligence 8.6, calibration 77.5, p50 0.756 s. Yes/no, choice and score, each with a reject option. CPU only: short states under 1 s, hard-tier median 4.3 s / p95 27 s.

  • Adoption2/100

    69 Hugging Face downloads and 2 likes, 20 GitHub stars, three Show HN posts by the author in three days (9, 1 and 3 points) and a handful of replies on X. No integrations or production use reported.

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.

Read this first

  1. One person, every number self-run. The JevBench 0.732 is on the public items with the author’s harness; there is no official JevBench or Decision Index result.
  2. CPU latency depends on length. Short states answer in under a second; the long policies in JevBench’s hard tier take a 4.3 s median and a 27 s p95 on the author’s machine. The card gives no p50 for short requests, so the frontmatter leaves latency blank.
  3. The decision is read from letter logits. The model is a fine-tuned chat LM that answers with one option letter; the runtime softmaxes the letter logits. The weights were trained for this, which is why it is in the catalog and not in (NOT)Models.

gutsy-0.8b is an open decision model published on Hugging Face on 1 October 2026 by an independent developer (kouhxp), with code on GitHub. The current release is v0.4 (2 October). It targets machines without a GPU: the reference file is a 775 MB Q8_0 GGUF served by llama.cpp, with a 505 MB Q4_K_M that the author says is 25–30% faster at the same panel scores.

Specs

Attribute Value
Base Qwen/Qwen3.5-0.8B
Training Full fine-tune, token embeddings frozen; cross-entropy + 0.5 × Brier on soft labels; ~105k questions, one epoch, one 80 GB GPU for about an hour
Decision readout Next-token logits of the option letters A–P (plus Z = “none of these”), softmaxed over those letters; temperatures fitted per type on validation
Question types yes/no, choice (up to 255 options via runtime shortlisting), score; reject probability on choice and score
Files gutsy-0.8b-v04-q8_0.gguf (775 MB), gutsy-0.8b-v04-q4_k_m.gguf (505 MB), calibration JSON per file
Serving gutsy-inference 0.4.0 (llama-cpp-python ≥ 0.3.35): POST /v1/systemone, POST /v1/decisions, /api/alpha/decisions; state caching
Context 8,192 tokens; English only
License Apache-2.0 (code and weights)
Launch (v0.4: 2 October)

Evidence (UnverifiedClaim)

From the README and model card, all self-run:

Metric v0.3 v0.4 (Q8_0) v0.4 (Q4_K_M)
JevBench public total (231) 167 (0.723) 169 (0.732) 161 (0.697)
Calibration error, easy / original / hard 0.023 / 0.122 / 0.078 0.022 / 0.107 / 0.137 0.019 / 0.137 / 0.125
BoolQ / SNLI / CommonsenseQA (700 each; Jev 0.906 / 0.883 / 0.863) 0.850 / 0.879 / 0.587 0.837 / 0.870 / 0.577 —
BANKING77, all 77 intents, zero-shot — 0.613 —

The author lists the weak spots plainly: date and number arithmetic (JevBench temporal_numeric 0.27), new multi-step rulebooks (long_policy 0.32), judging close answer pairs (JudgeBench 0.54, down from 0.60 in v0.3), world knowledge, and worse calibration on hard, unfamiliar items than v0.3.

Fit / anti-fit

Fit: local, private yes/no and routing checks on a laptop or small server with no GPU; prototypes that already speak /v1/systemone; short states.

Anti-fit: arithmetic, dates and long policies; knowledge-heavy questions; anything latency-sensitive with long inputs on CPU; non-English text.

Why it is in the catalog

gutsy fine-tunes all of Qwen3.5-0.8B’s weights for typed decisions (with a Brier term and fitted temperatures), publishes them under Apache-2.0 with a documented typed API, and covers all three decision types. Reading the answer from letter logits is how the decision is decoded, not a sign of a frozen model: the (NOT)Models line is “no decision training”, and gutsy has it.

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) Gutsy 0.8B v0.3 Capability 43.1; Intelligence 8.6; calibration 77.5; p50 0.756 s; $0.0053/1k

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

  • “One person, every number self-run. The JevBench 0.732 is on the public items with the author’s harness; there is no official JevBench or Decision Inde…”

Score update (audit, 17 Oct 2026). Capability 6.5 → 5.7 = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 10 × 0.30 + 7 × 0.25 + 3 × 0.25 + 1 × 0.20 = 5.70 → 5.7. Types, calibration and latency come from the earlier working (10 / 7 / 3); only accuracy is re-derived from the board (1; was 5): JevBench Intelligence 8.6 ÷ 63.6 × 8 = 1.1, half-up. Maturity and adoption unchanged.

Hugging Face on 17 Oct (kouhxp/gutsy): 447 downloads, 2 likes, licence tag apache-2.0, last modified 2026-10-05.