In short
- Kev is Jared Palmer's family of text decision models, on Ollama from 17 October 2026 — it scores answers to structured questions in a single forward pass and never generates text.
- Four checkpoints (0.8B, 4B, 9B, 27B); tags run from 812MB to 18GB, with kev:latest at 5.6–6.7GB. Budget roughly the download size in memory plus context headroom.
- The default is kev:9b-nvfp4: 68/111 hard JevBench tasks at 11.16 GiB resident — better than 9B BF16 on the hard subset while using less memory.
- Requires Ollama 0.40.4+ and the Apple Silicon MLX runner, so it's Mac-only for now; we've found no other engine instructions as of release day.
- All quality numbers are the publisher's own runs on one public JevBench set — test on your own labelled examples before trusting thresholds.
What kev is
Kev is a family of text decision models from Jared Palmer, and as of today (17 October 2026) it's on the Ollama library. If you've been following this space for the past fortnight, the shape of it will feel familiar: it's the latest in a small wave of models that answer with probabilities instead of prose — we've covered Strands Decider, nimble-v3 and Clef Flash in recent weeks, and the readme itself names the Tev1, Nimble and Clef Flash pattern of a capable compact default plus explicit larger and source-precision tags.
It doesn't chat. You give it a state — a supplied text, such as a support message or a document excerpt — plus one or more typed questions. In a single forward pass it scores the answers and returns a choice, probabilities for each option, and a confidence value. No prose to parse, no tokens to wait for.
It doesn't generate text or accept images. It requires Ollama 0.40.4 or later, and the tags use the Apple Silicon MLX runner — so this is a Mac-only release for now.
Sizes, memory and which hardware suits
There are four publisher checkpoints — 0.8B, 4B, 9B and 27B — and the Ollama tags range from 812MB (0.8b-q8_0) to 18GB (9b-bf16). kev:latest sits at 5.6–6.7GB. One thing worth making explicit: pulling kev gets you kev:9b-nvfp4 — the readme's default — so that 5.6–6.7GB download is the compressed 9B checkpoint. The readme also lists 27B tags (BF16 and NVFP4) with their measured memory footprints, though the download sizes shown on the library page top out at the 9B tags.
Our usual rule of thumb applies: budget roughly the download size in RAM or VRAM, plus headroom for context. The tag's num_ctx is 73,728 tokens — but the readme is explicit that a large accepted input doesn't imply reliable accuracy at that length, so don't treat the advertised 256K context as a quality guarantee.
The readme also gives observed resident sizes after a three-question workload on a 128GB M5 Max — useful for calibration, though not peak requirements:
| Tag | Format | Resident memory |
|---|---|---|
| kev:0.8b-mlx-bf16 | BF16 | 3.28 GiB |
| kev:0.8b-mxfp8 | MXFP8 | 2.84 GiB |
| kev:4b-mlx-bf16 | BF16 | 12.47 GiB |
| kev:9b-mlx-bf16 | BF16 | 19.59 GiB |
| kev:9b-nvfp4 | NVFP4/MXFP8 | 11.16 GiB |
| kev:27b-mlx-bf16 | BF16 | 55.56 GiB |
| kev:27b-nvfp4 | NVFP4/MXFP8 | 25.56 GiB |
In hardware terms, working from those resident figures plus context headroom: the 0.8B tags (812MB–1.5GB downloads, ~3 GiB resident) are the only ones that plausibly fit a small machine; 4B BF16 at 12.47 GiB resident wants a comfortable 16GB-class Mac; the 9b-nvfp4 default at 11.16 GiB sits in the same bracket; 27b-nvfp4 at 25.56 GiB points at 32GB+ machines; and 27B BF16 at 55.56 GiB is firmly in big-unified-memory territory. These are our inferences from the readme's measurements, not the publisher's stated hardware requirements.
Because these are MLX tags, NVIDIA GPU owners are out of luck until other formats appear. Worth saying plainly: this is a Mac release.
The readme's own pick is kev:9b-nvfp4 — the best balance of quality, response time and memory on the tested M5 Max.
How to run it
Pull it as usual — you'll need Ollama 0.40.4 or later:
ollama pull kev
You don't chat with it. You send a System One request to the local server with a model, a state, and typed questions. The readme's routing example:
curl http://localhost:11434/v1/systemone \
-H 'Content-Type: application/json' \
-d '{
"model": "kev",
"state": "The package arrived with a cracked screen.",
"questions": {
"route": {
"type": "choice",
"instructions": "Which team should handle this message?",
"criteria": {
"returns": "Damaged items and returns",
"shipping": "Delivery and tracking",
"billing": "Payments and charges"
}
}
}
}'
The response contains a choice, probabilities for each option, and a confidence value, rounded to four decimal places. One caveat worth flagging: the readme says confidence describes how concentrated the choices are, not the chance the answer is correct. Don't read it as a correctness score.
Three question types are supported:
- choice — pick from 1–255 named options.
- noul — the probability that something is true.
- score — an expected level over ordered criteria, with 1–255 levels supported.
Multiple questions can share one state in a single request, which is where the one-forward-pass efficiency shows up: one state, several decisions, one call.
As of today, the Ollama library page is the only documentation we've seen — the same situation Clef Flash was in on its release day. We've found no LM Studio, llama.cpp or vLLM instructions, so don't go hunting for Hugging Face pages that aren't linked. If you've set up another decision model through this System One API before — nimble-v3 uses the identical endpoint and Ollama version requirement — the request shape above should look familiar.
The numbers that matter: JevBench and the NVFP4 surprise
The readme publishes JevBench v1.2.16 results — 231 public tasks, final-runtime local results, with 303 sealed or judge tasks excluded. It's not an official leaderboard score, and it's one public set rather than a guarantee for your task.
| Tag | Public correct | Hard subset | Resident memory |
|---|---|---|---|
| 0.8b BF16 / MXFP8 | 146/231 | 39/111 | 3.28 / 2.84 GiB |
| 4b BF16 | 175/231 | 60/111 | 12.47 GiB |
| 9b BF16 | 184/231 | 66/111 | 19.59 GiB |
| 9b-nvfp4 | 183/231 | 68/111 | 11.16 GiB |
| 27b BF16 | 202/231 | 83/111 | 55.56 GiB |
| 27b-nvfp4 | 205/231 | 86/111 | 25.56 GiB |
The headline is the 9B NVFP4 row: it answered 68 of 111 hard tasks while using less observed resident memory (11.16 GiB) than 4B BF16 (60/111 at 12.47 GiB) or 9B BF16 (66/111 at 19.59 GiB). Better hard-subset accuracy than the full-precision 9B, at roughly half the memory. The 27B NVFP4 tag goes further: 86/111 hard tasks at 25.56 GiB resident.
Two honest notes from the readme itself. First, quantisation changed some individual answers and probabilities even where aggregate accuracy was close — the publisher says lower-memory formats are additional choices, not replacements for the BF16 checkpoints. Second, these are results from one public set. If your task looks nothing like JevBench, none of these numbers transfer automatically.
The usual release-day caveat applies, the same one we gave for Bonsai 2: every quality number here is the publisher's own run, and we haven't seen any independent benchmarks yet.
What it's good for, and where it falls over
Kev is built for bounded decisions where the evidence and permitted answers are supplied: routing a message to a team, yes/no checks, ordered scoring. One forward pass instead of generating and parsing prose, with probabilities you can threshold. That's a genuinely useful primitive for local pipelines — triage, guards, classification.
The limits are the publisher's own, and they're worth taking seriously:
- Accuracy falls on unfamiliar domains, long inputs, reordered options, and tasks needing facts absent from the state.
- Evaluations focus on English.
- The 0.8B models got only 39/111 hard tasks — simple decisions only.
- A fitted temperature doesn't make probabilities universally calibrated, and the readme says plainly: don't use these as the sole control for consequential decisions.
The practical advice from the readme is sound: add an explicit none option when no listed answer may fit, and check output quality and probability thresholds on labelled examples from your own task — especially after changing size or quantisation.
If you need a general-purpose classifier or guard today, with broader engine support, llama-guard3:1b and granite3-guardian:2b are established alternatives. They answer differently — generated labels rather than scored choices — but they run in more engines and have far more independent usage behind them.
What we'd pick: on a Mac with room for ~11 GiB resident plus context, kev:9b-nvfp4 — it's the readme's default and the numbers justify it. On a machine with ~26 GiB to spare, 27b-nvfp4 if response time is acceptable. We'd skip the 0.8B tags for anything beyond the simplest routing.
Worth hosting?
Decision models are cheap to serve per request — one forward pass, no long generations — which makes this category a good fit for hardware that sits idle. A Mac that comfortably holds the 9b-nvfp4 default can serve routing or classification requests all day.
Whether kev itself can be served through a marketplace like ours depends on it being listed — as of today it isn't on ours. But the general point stands: the connector supports Ollama, so an idle Mac running open models can earn from spare capacity. Hosts keep everything except the 15% platform fee, and new hosts get their first 12 months fee-free (up to £250 of earnings) — details on the providers page.
Questions people ask
What is kev on Ollama?
Kev is a family of text decision models from Jared Palmer, added to the Ollama library on 17 October 2026. It reads a supplied text state plus typed questions and returns choices, probabilities and confidence in one forward pass — it never generates text.
How much RAM does kev need?
Budget roughly the tag's download size in memory, plus context headroom. The readme's measured resident sizes run from 2.84 GiB (0.8B MXFP8) to 55.56 GiB (27B BF16); the default 9b-nvfp4 sat at 11.16 GiB on the test machine.
How do I run kev locally?
Install Ollama 0.40.4 or later, then ollama pull kev and send a System One request to http://localhost:11434/v1/systemone with a model, a state and typed questions.
Is kev a chat model?
No. It doesn't generate text or accept images. It scores answers to structured questions — choice, noul (probability of true) and score (ordered levels) — and returns probabilities instead of prose.
Does kev work on Windows or NVIDIA GPUs?
Not yet. The tags use the Apple Silicon MLX runner, so as of 17 October 2026 kev is Mac-only.
Which kev tag should I pick?
The readme picks kev:9b-nvfp4 as the default — the best balance of quality, response time and memory. The 27B variants are the quality choices; the 0.8B variants suit tight memory budgets and simple decisions only.
Models in this post — live on Inpherio
Have a GPU, a Mac or a spare PC?
Let it earn while you sleep.
Run open models for people who need them. You set the price and keep everything but the 15% fee.
Become a hostBuilding with AI?
Every open model, one API.
OpenAI-compatible. Priced by the people who run the hardware.
Start free See the marketplaceRead “Kev on Ollama: Jared Palmer's decision model family, hands on” at https://inpherio.co.uk/blog/kev-on-ollama-jared-palmer-s-decision-model-family-hands-on