Model tracker / Decision models / Jared Palmer
Kev-27B
Largest open-weight Kev decision model, fully fine-tuned Qwen3.8-27B; needs an 80 GB GPU
Kev-27B is an open-weight decision model from Jared Palmer (Apache 2.0, plus training code and eval suites, 27B, fully fine-tuned from Qwen3.8-27B (51 GB of weights) parameters): it answers typed questions, such as yes or no, pick one or pick several, with probabilities instead of writing text. There's no hosted version: you run it on your own hardware. It scores 78.9 (self-hosted; seconds are the board's estimate; not capability-ranked, as its cost is above the board's cap) on the independent JevBench Capability Score (board v1.7.46), where Jev, the reference, scores 77.1.
- Made by
- Jared Palmer
- Weights
- Open, Apache 2.0, plus training code and eval suites
- Size
- 27B, fully fine-tuned from Qwen3.8-27B (51 GB of weights)
- Where
- Self-host only
- Images
- No, text only
- Context
- 65,536
- Fine-tuning
- Full recipe and a fine-tuning skill for coding agents published
Kev-27B price
US dollars per million input tokens. Decision models don't write text, so none of the hosted ones charge for output.
| Where | Input | Output | Notes | Checked |
|---|---|---|---|---|
| Self-host only | Self-host | – | – | 2026-10-25 |
Kev-27B benchmarks
| Benchmark | Result | What it measures | Reported by |
|---|---|---|---|
| JevBench v1.7.46 | 78.9 | Capability Score: accuracy and calibration on typed decisions, with Jev as the reference system | Independent |
| JevBench time | 0.52 s | Seconds per decision: measured end to end for hosted APIs, the board's estimate for self-hosted models | Independent |
| Reported speed | 67.2 ms model time for six questions on an H200 | Latency on the maker's own hardware | Jared Palmer |
Makers' speeds are measured on their own hardware and aren't comparable with each other or with JevBench's timings.
Decision models like Kev-27B
Every other decision model we track, by independent JevBench score.
| Model | Input / M | JevBench | Weights |
|---|---|---|---|
| Kev-27B | Self-host | 78.9 | Open |
| Mammoth-1 Baseten (Base Labs) | Free | 85.2 | Open |
| Perplexity Decider V1.2 Perplexity | $0.02 | 80.8 | Open |
| d1-ultra Liquid AI | $0.042 | 80.5 | Closed |
| Clef Cloudflare | $0.24 | 75.3 | Open |
| Jev 1.14 TypeSafe AI | $0.042 | 75.0 | Closed |
| d1 Liquid AI | $0.02 | 74.4 | Closed |
| Mercury Decide Inception | $0.02 | 73.6 | Closed |
| OpenAI Decisions (GPT-6 Luna) OpenAI | $0.10 | 73.5 | Closed |
| Microsoft-Decision-1 Microsoft | $0.042 | 70.8 | Closed |
| Elephant-1 Baseten (Base Labs) | Free | 70.2 | Open |
| Strands Decider 26B-A4B AWS Strands Labs | Self-host | 66.7 | Open |
| Clef-flash Cloudflare | $0.038 | 63.8 | Open |
| Perplexity Decider Flash V1.2 Perplexity | $0.01 | 61.7 | Open |
| Clef-omni Cloudflare | $0.15 | 58.3 | Open |
| d1-3B Liquid AI | Self-host | 49.5 | Open |
| d1-omni-600M (experimental) Liquid AI | Self-host | 32.8 | Open |
| Laya Convai Innovations | Free | 32.5 | Open |
| CLM-8B Stanford and NVIDIA | Self-host | 20.1 | Open |
| Clairic Cloudflare | $0.10 | – | Closed |
| Kev-0.8B Jared Palmer | Self-host | – | Open |
| Kev-4B Jared Palmer | Self-host | – | Open |
| Kev-9B Jared Palmer | Self-host | – | Open |
| Strands Decider 2B AWS Strands Labs | Self-host | – | Open |
Compare them all in the decision models table →
Kev-27B FAQ
How much does Kev-27B cost?
Kev-27B has no hosted, per-token price: you download it and run it yourself, so the cost is your own compute.
Is Kev-27B open weight?
Yes. Jared Palmer publishes the weights under Apache 2.0, plus training code and eval suites, so you can run and fine-tune it yourself.
How fast is Kev-27B?
Jared Palmer reports 67.2 ms model time for six questions on an H200, measured on its own hardware. JevBench gives 0.52 seconds per decision, its estimate for self-hosting.
How does Kev-27B compare with Jev?
On JevBench's Capability Score (board v1.7.46), Kev-27B scores 78.9 against Jev's 77.1. Jev costs $0.042 per million input tokens; Kev-27B is self-host only. JevBench is a one-person project that uses Jev as its reference, which favours cheap hosted models.