Ninety-five dollars of rented H100 time produced Kev's three-model Qwen3.5 port
Jared Palmer's Kev now ships as 0.8B, 4B and 9B variants that score typed answer options over frozen Qwen3.5 weights, and on his own development comparison hosted Jev still scores higher than the largest of them.
The Engineer · Build desk
What happened
- Jared Palmer released a Qwen3.5-based generation of Kev on September 20th, making the open-source decision-model project a three-model family developers can train, inspect and run on their own hardware.
- Kev takes a document or other application state, evaluates a set of typed questions, and returns probabilities for yes-or-no, multiple-choice and ordered-score answers.
- Palmer's execution log puts the Qwen3.5 work at roughly $95 of Modal H100 usage plus $0.03 of Jev API calls, covering probes, several training trials, benchmark runs, locked test reads and an ablation.
- On the repository's locked, new-source test set, Kev-9B scored 0.837 accuracy against 0.780 for the prior Kev-8B, with Kev-4B at 0.832 and the 0.8B variant at 0.668.
- The code and model adapters ship under Apache 2.0 alongside frozen evaluation data and model cards, and Kev can be called through TypeSafe's Python SDK with the server and weights local.
Why it matters
- What doesn't add up contradiction The repository's own development comparison puts hosted Jev at 0.857 and Kev-9B at 0.812, and Palmer warns about that comparison, so a team choosing self-hosting is doing it on a number the author himself flags.
- What it limits constraint Kev never generates explanatory text at inference, so any workflow that needs a written rationale next to the decision has to produce it with a second model or from the probabilities.
- What it costs cost The logged spend bought the Qwen3.5 port experiments, and Kev's development predates that port, so budgeting a decision-model family from the $95 figure budgets for a port and not for the family.
- Who has to decide decision Nobody outside the project has reproduced these scores. Whoever adopts Kev carries the evaluation work: the released harness and frozen data are the only way to check the claim against your own tickets.
Next story · Build · Oct 8
Benchmark and Index put $60 million behind Hone's long-running business agentsConfirmed2 publishers
A call into Kev scores options. Each released checkpoint is a rank-16 LoRA adapter and a small pointer-style readout head attached to a Qwen base model. The head scores the answer options supplied in the request, and a softmax turns those scores into probabilities [10]. None of that inference time goes into explanatory prose [11]. The base weights stay fixed while the adapter and the head learn the decision task [21].
The port cost Palmer his batching trick. Qwen3.5's bases mix conventional attention with recurrent Gated DeltaNet layers, and those layers do not respect the block attention mask the earlier Qwen3 version relied on [26]. Kev now runs each question in an independent row, computing the shared state once and reusing the cache, so questions stay isolated and the input is not re-encoded per question [27].
Training ran two epochs over 12,576 examples: 10,000 from ten public datasets, 896 generated policy examples and 1,680 built from 60 generated rule structures [21][24]. Palmer built Kev with help from Devin, the coding agent made by Cognition, where he is vice president of engineering [18]. His repository says no Jev outputs were used in training [19]. Palmer's execution log itemizes spend down to three cents of API calls, more bookkeeping than most model releases publish [20].
The accuracy table is a claim about Palmer's questions. For it to transfer, your typed questions would have to resemble ten public datasets plus generated policy and rule examples. Your definition of a correct answer would have to match the harness that scored them [21]. Kev-4B sits 0.005 behind the 9B on the locked set [22].
Where Palmer compared the two systems directly, on the repository's development comparison, hosted Jev leads Kev-9B by 0.045 [23], and Palmer explicitly warns about that comparison [17]. The same 9B checkpoint also scores 0.025 higher on the locked, new-source set than it does on the development comparison [25]. The generation-over-generation gain and the gap to hosted Jev come from different questions.
TypeSafe AI introduced Jev and its "System One" category on September 15th, and founder Diogo Almeida described it as a model for fast, typed decisions that software can consume directly, according to runtimewire [5][6]. Kev answers through TypeSafe's Python SDK while the server, weights and training code stay local [7].
The Qwen3.5 family landed five days after the Jev announcement, though Kev's development predates the port [9]. Code and adapters are Apache 2.0, with frozen evaluation data and model cards published alongside [8]. A team that doubts the numbers can rerun the harness that produced them on its own tickets.
What to watch
- An independent reproduction of Kev's locked test-set scores on data the project did not curate.
- Any change to TypeSafe's Python SDK or its terms that affects clients pointing at a self-hosted Kev server.
- A published cost log for Kev's pre-Qwen3.5 development, which would show what the family cost end to end.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption
- Not scored
- Hype gap+10
- Incentives35
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Jared Palmer released a Qwen3.5-based generation of Kev on September 20th, turning the open-source decision-model project into a three-model family that developers can train, inspect and run on their own hardware.
- [2]
The Kev release includes 0.8B, 4B and 9B parameter variants.
- [3]
Kev takes a document or other application state, evaluates a set of typed questions and returns probabilities for yes-or-no, multiple-choice and ordered-score answers.
- [4]
Kev is designed for the small semantic judgments buried inside software workflows: routing a support ticket, choosing an escalation path or scoring a customer's frustration.
- [5]
TypeSafe AI introduced Jev and its "System One" model category on September 15th.
- [6]
TypeSafe founder Diogo Almeida described Jev as a model for fast, typed decisions that software can consume directly; instead of generating a paragraph for an application to parse, Jev returns values and probabilities from a schema supplied in advance.
- [7]
Kev follows Jev's interface and can be called through TypeSafe's Python SDK, while keeping the server, weights and training code local.
- [8]
Kev's code and model adapters are available under Apache 2.0, along with frozen evaluation data and model cards.
- [9]
The Qwen3.5 Kev family arrived five days after TypeSafe's public Jev announcement, although Kev's development predates the port.
- [10]
Each Kev checkpoint consists of a rank-16 LoRA adapter and a small pointer-style readout head attached to a Qwen base model; the model scores the supplied answer options directly and applies a softmax to produce probabilities.
- [11]
Kev does not spend inference time generating explanatory prose.
- [12]
On the repository's locked, new-source test set, Kev-9B scored 0.837 accuracy compared with 0.780 for the prior Kev-8B.
- [13]
Kev-4B reached 0.832 on the locked test set, up from 0.806 for its Qwen3 predecessor.
- [14]
The 0.8B Kev model reached 0.668 on the locked test set, against 0.642 for the earlier 0.6B checkpoint.
- [15]
The accuracy figures come from Kev's own evaluation harness and have not been independently reproduced.
- [16]
Hosted Jev scored higher on the repository's development comparison, at 0.857 accuracy against 0.812 for Kev-9B.
- [17]
Palmer explicitly warns about the development comparison in which hosted Jev scored 0.857 against Kev-9B's 0.812.
- [18]
Palmer built Kev with assistance from Devin, the coding agent made by Cognition, where he is vice president of engineering.
- [19]
Palmer's repository says no Jev outputs were used in training Kev.
- [20]
Palmer's execution log records roughly $95 of Modal H100 usage for the Qwen3.5 experiments, plus $0.03 in Jev API calls, covering probes, multiple training trials, benchmark runs, locked test reads and an ablation.
- [21]
The released Kev models were trained for two epochs on 10,000 examples from ten public datasets, 896 generated policy examples and 1,680 examples built from 60 generated rule structures, with base weights fixed while the adapter and pointer head learn the decision task.
- [22]
Kev-4B trails Kev-9B by 0.005 accuracy on the locked, new-source test set.
- [23]
On the repository's development comparison, hosted Jev leads Kev-9B by 0.045 accuracy.
- [24]
The released Kev models were trained on 12,576 examples in total.
- [25]
Kev-9B scores 0.025 higher on the locked, new-source test set than on the development comparison.
- [26]
The Qwen3.5 base models combine conventional attention with recurrent Gated DeltaNet layers, which do not respect the block attention mask used in the earlier Qwen3 version of Kev.
- [27]
Kev now runs each question in an independent row, computes the shared state once and reuses its cache, preserving isolation between questions while avoiding repeated work on the input.
Sources
2 independent publishers whose own reporting we read for this story.
- runtimewire.comJared Palmer ports Kev to Qwen3.5 for roughly $95 in H100 time
1 article · September 21, 2026
- thenewstack.ioYour AI agent is burning tokens on choices that don’t need words
1 article · September 21, 2026
Topics and entities
Topics
- AI Agents
- Open-Weight Model Publishing
- LoRA Fine-Tuning
- Typed decision models