Kev
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity6.4/10
Kev 1.0: Apache-2.0 weights in four sizes, GPU and MLX server, browser demo, fine-tune / deploy skills, cards with data sources. Kev-4B is on OpenRouter and, since 2 Oct, in llama.cpp's /v1/systemone (ggml-org GGUF). TypeSafe SDK works by changing endpoint. No SLA.
- Capability8.1/10
Official boards (17 Oct): Decision Index 0.3.2 Full 58.78 (#24); JevBench Intelligence 64.2, calibration 93.7, p50 0.180 s. Noul / Choice / Score with a pointer head and no generated tokens. Temperature-fitted, pooled ECE 0.019 for Kev-27B (author). Median model time 23–67 ms on GPU (author, no p99).
- Adoption26/100
The most-starred open Jev reconstruction (~8.2k GitHub stars, 520+ forks), tens of thousands of HF downloads, an OpenRouter listing and derivatives such as NeoHorse-Jev. Kev 1.0 launch post: ~110 likes in its first hours. No named production users.
Vendor claims
- Generalization index (14 public datasets excluded from fine-tuning, chance-corrected): Kev-27B 52.3, Kev-9B 41.0, Kev-4B 38.0, Kev-0.8B 23.3[Vendor claim — not independently verified]
- Kev-27B pooled calibration error 0.019 on the 14-dataset test; 0.073 when averaged per dataset[Vendor claim — not independently verified]
- Median model time on Modal: Kev-27B 67 ms (H200), Kev-9B 24 ms, Kev-4B 42 ms, Kev-0.8B 23 ms; Kev-4B on a 32 GB M5: 721 ms for new text, 136 ms cached[Vendor claim — not independently verified]
- Kev-27B accuracy: 75.7% generalization, 91.8% policy and rule reasoning, 90.8% consumer complaints, 79.0% developer-tool decisions (author's harness)[Vendor claim — not independently verified]
News covering this model
- Taktile Labs benchmark: a decision model fine-tuned on 64 bank statements beats every out-of-the-box model
- Paper: decision fine-tuning does not make a better reranker (Kev, Clef vs reranker fine-tunes)
- Kev 1.0: Jared Palmer's open decision family gets a full-weight 27B
- Kev 4B lands on OpenRouter — first open System One model on a major gateway
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.
Kev is an open reconstruction of the System One / Jev contract by Jared Palmer. It started on 18 September as a LoRA adapter + readout head on Qwen2.5-0.5B; with Kev 1.0 it is a family of four models from 0.8B to 27B.
It reads a document once and answers many typed questions in parallel in a single prefill pass — no autoregressive decode. Question types: noul (yes/no → catalog boolean), choice, and score. Output is a calibrated distribution over options, not prose.
Kev 1.0 (1 October 2026)
Jared Palmer released Kev 1.0 on 1 October 2026 (post on X, 20:33 BRT): four open decision models from 0.8B to 27B, all Apache-2.0, with a new Kev-27B and an updated Kev-9B. The API is the same as TypeSafe’s, so TypeSafe SDK users switch by changing the endpoint and model name. The server runs on GPUs and on Apple Silicon through MLX; agent skills cover fine-tuning and deployment.
| Model | What changed | Runs on | Median model time |
|---|---|---|---|
| Kev-27B | New: every weight of the Qwen3.8-27B text backbone fine-tuned (smaller Kevs keep LoRA on Qwen3.5); ~146k examples / 337k questions, training documents up to 32k tokens, 64k-token document window; released backbone blends 85% new fine-tune with 15% of the previous 27B | H200 (or H100, B200) | 67 ms |
| Kev-9B | One more fine-tuning pass; policy reasoning 58% → 83% (author) | H100 (or L40S) | 24 ms* |
| Kev-4B | Unchanged | L40S or a 32 GB Mac | 42 ms |
| Kev-0.8B | Unchanged | L4 or any Apple Silicon Mac | 23 ms |
* Measured on the previous 9B. All numbers are from the author’s harness (UnverifiedClaim). The author says no Jev outputs were used in training.
Caveats from the post: the new 27B is less accurate and more overconfident on long legal contracts than the previous one (still available as kev-27b@v1-lora); base-model exposure to the test data is unknown; checkpoint selection was not fully blind. On the earlier weights, the official Decision Index gives Kev 9B 38.48, Kev 4B 34.64 and Kev 0.8B 14.6; Kev 1.0 is not measured there yet.
Catalog note: the page keeps
artifactKind: lora-adapterbecause three of the four sizes are adapters; Kev-27B is now a full fine-tune.
What it is
- Open weights (Kev 1.0):
kev-27b,kev-9b,kev-4b,kev-0.8b(Apache-2.0; base under Qwen license). Earlier:kev-0.5b,kev-0.6b - Repo + local server: github.com/jaredpalmer/kev
- API:
POST /v1/systemone— TypeSafe-compatible shapes; official SDK works with a localbase_url - Architecture follows the independent reconstruction in Archer Hume — Jev’s Architecture Unmasked (17 Sep 2026): shared state, isolated question branches, pointer readout, proper scoring — not a TypeSafe confirmation
See also the catalog entry for the closed reference product: Jev.
Distribution
- OpenRouter (NEW):
jaredpalmer/kev-4b— $0.01 / $0 per 1M tokens (in/out), see news. - Local server: github.com/jaredpalmer/kev — run with your own GPU
JevBench v1.4.2 (pre-1.0 Kev family)
| Variant | Overall | Intelligence | Calibration | Speed | Cost | Rank |
|---|---|---|---|---|---|---|
| Kev 4B | 36.1 | 42.1 | 68.4 | 74.2 | 58.3 | #24 |
| Kev 0.6B | 29.1 | 37.2 | 62.1 | 79.4 | 75.2 | #37 |
| Kev 0.5B | 28.4 | 36.8 | 61.2 | 78.9 | 76.8 | #39 |
The benchmark confirms Jared Palmer’s self-reported ~-19pp transfer gap vs Jev. The 4B variant closed the gap somewhat, but before 1.0 Kev was a prototype / research tool, not a production peer.
Derivatives
- NeoHorse-Jev-4B (TokenRhythm, 23 Sep 2026) reuses Kev’s inference code (
model.py/schema.py, credited) and pointer-head design on the NeoHorse-1-4B backbone. Its head is its own training: none of its tensors matchkev-4b’shead.pt. Official Decision Index 0.2.1: 36.75 (#30) vs Kev 4B 34.64 (#32).
In use: LocalFirstAI answerability gate (19 Oct 2026)
publicado (OK do Fabricio 19/10 BRT).
LocalFirstAI (MIT; a one-person lab running local models on a Mac Mini M4 Pro) tested four local categorizers as a gate that decides, before a 4B model answers, whether the retrieved material can answer the question. In exp_042 (scored 18 Oct, 150 sealed questions reused from an earlier run), Clef 27B was the primary gate. Kev-4B came close: refusals on the 38 should-refuse questions rose from 0.474 to 0.921, exact match from 0.753 to 0.820, with 1 false refusal among the 101 answerable ones, at about 2 s per question against about 12 s for Clef 27B (served by llama.cpp llama-server, about 4.5 GB).
The author then wired Kev-4B into the project’s answering pipeline and, since 18 Oct, runs it in shadow mode in a live guide bot: it logs a probability and blocks nothing. A first replay showed the thresholds do not transfer (it would have refused 29 of 60 answerable questions), so it stays in shadow mode. A separate test (exp_043) used Kev-4B alone to triage 235 synthetic estate-agency leads: intent 0.894, spam caught 29 of 32 with 3 of 203 real leads wrongly flagged. These are third-party numbers we have not reproduced. Scores unchanged; it is a named user, but not production traffic yet. Found through a TeksEdge post (19 Oct).
Limits (read before you ship)
- Pre-1.0 README label: research prototype — “not a production model and it is not Jev”
- Pre-1.0 JevBench runs confirmed a ~−19pp transfer gap vs Jev for the 0.5B (~−14pp for the 4B); Kev 1.0 has no third-party run yet
- Residual option-order sensitivity (~7% argmax flips in author tests)
- Local serve, no auth — fine for laptop demos, not a hosted waitlist SLA
- Do not use for production decisions that affect people without your own labelled eval
Why it is in the catalog
Public typed decision I/O, documented System One-compatible API, and open weights. It sits next to Jev as an open / local point on the quadrant. Since Kev 1.0 the point represents the family, led by Kev-27B.
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 |
|---|---|---|
| Decision Index 0.3.2 (16 Oct 19:00 UTC) | Kev 27B | Full 58.78 (#24, tie group 7); public 56.69 (#49); same-skills 55.51; new-domains 55.49 |
| Decision Index 0.3.2 (16 Oct 19:00 UTC) | Kev 9B v2 | Full 43.27 (#86, tie group 20); public 41.54 (#92); same-skills 42.37; new-domains 41.03 |
| Decision Index 0.3.2 (16 Oct 19:00 UTC) | Kev 4B r10 | Full 39.5 (#98, tie group 23); public 37.95 (#106); same-skills 36.17; new-domains 41.25 |
| Decision Index 0.3.2 (16 Oct 19:00 UTC) | Kev 0.8B r15 | Full 13.66 (#147, tie group 44); public 16.78 (#153); same-skills 15.84; new-domains 13.2 |
| JevBench open weights (live 17 Oct) | Kev 27B | Capability 78.9; Intelligence 64.2; calibration 93.7; p50 0.180 s; $0.53/1k |
| JevBench open weights (live 17 Oct) | kev 8B | Capability 50.0; Intelligence 29.3; calibration 70.8; p50 0.074 s; $0.097/1k |
| JevBench open weights (live 17 Oct) | kev 4B | Capability 44.0; Intelligence 19.5; calibration 68.5; p50 0.072 s; $0.014/1k |
| JevBench open weights (live 17 Oct) | kev 0.6B | Capability 36.9; Intelligence 7.6; calibration 66.3; p50 0.057 s; $0.0046/1k |
| JevBench open weights (live 17 Oct) | kev 0.5B | Capability 32.9; Intelligence 4.6; calibration 61.2; p50 0.044 s; $0.0046/1k |
Board-status statements on this page to recheck against the table (may be superseded):
- “Official Decision Index 0.2.1: 36.75 (#30) vs Kev 4B 34.64 (#32).”
Score update (audit, 17 Oct 2026). Capability 7.7 → 8.1 = 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 6): Decision Index Full 58.78 ÷ 60.11 × 8 = 7.8, half-up. Maturity and adoption unchanged.
Hugging Face on 17 Oct (jaredpalmer/kev-27b): 2,619 downloads, 41 likes, licence tag apache-2.0, last modified 2026-10-01.
Links
- llama.cpp:
llama serve -hf ggml-org/Kev-4B-GGUF, thenPOST /v1/systemone(ggml-org blog, 2 Oct 2026)