TypeLLM
@TypeLLM

LLMs with type-safe generation

 London
Based in United Kingdom
typellm.ai
Joined December 2025
Jev answers every question instantly. That's great, until a question is hard. TypeLLM is System 1 when it can be, System 2 when it has to be. "thinking": "auto" picks per question, per call. Read more: typellm.ai/blog/thinking-aut…
1
10
673
Thinking in TypeLLM is now 3x faster, with no loss in accuracy. Get better typed decisions with faster thinking Try it now: typellm.ai/dashboard/playgro…
1
26
1,811
In Jev, every field is generated independently. In TypeLLM, earlier answers decide which fields come next: classify an email, then extract the amount if it is an invoice, or the date if it is a meeting request. All in one API call, with `when`.
8
4
86
7,107
System 1 models like Jev answer at once: fast, cheap, right when the answer is obvious. System 2 models reason step by step first: slower, but right when the steps matter. Most models are one or the other. TypeLLM decides per field, on every call.
1
3
254
How many r's are in "strawberry"? System 1 models like Jev answer in one shot and get it wrong. TypeLLM thinks first and says 3, and still answers "hello" (0) instantly. Two systems, one model. 🧵
1
21
2,971
Not another Jev. TypeLLM adds type-safe generation to LLMs — more types (string, number, boolean, enum), vision, dependent fields and thinking. Playground is open to all with $5 credit. API is rolling out to early-access users in the coming days. typellm.ai/dashboard/playgro…
13
65
528
163,033
If Qwen favours the first position, let every answer take a turn there—then average the resulting probabilities. Averaging just 8 permutations reduced the KL error by ~79%. Averaging all 720 reduced it by ~97%. Jev improved only slightly.
2
3
2,519
I asked Jev and TypeLLM to simulate rolling a fair die. TypeLLM gives reasonable results. Jev keeps giving me 1. Why? Here’s what we found 🧵👇
11
13
153
37,786
Benchmark
1
31
7,301
Jev can’t do any of these: 1. Image inputs 2. String, integer, and number types 3. Dependency execution in a single call TypeLLM can. Try it: github.com/TypeLLM/TypeLLM
46
264
2,853
195,329
Something Jev from @typesafeai — and other similar fine-tuned Jev models — can’t easily support: native numerical outputs. TypeLLM can directly generate constrained integers and floats, while remaining type-safe.
3
1,315
Jev @typesafeai, but with thinking? TypeLLM can think before producing a type-safe output. It dramatically boosts accuracy without any fine-tuning—surpassing Jev and GPT-5.6 Luna, and coming close to GPT-6 Astra! Open-source: github.com/TypeLLM/TypeLLM
3
3
24
3,895
Replying to @melvynx
But typeLLM can github.com/TypeLLM/TypeLLM
7
Replying to @hbouammar
Or one can enable traditional LLM doing the same type-safe generation, checkout typeLLM: github.com/TypeLLM/TypeLLM
8
Replying to @altryne
You can already do that with the image input in github.com/TypeLLM/TypeLLM
2
226
Replying to @moritzkremb
Images are supported in TypeLLM: github.com/TypeAR-AI/TypeAR
6
Replying to @HamelHusain
It's here! For regression, you just need to support float output type, checkout TypeLLM: github.com/TypeLLM/TypeLLM
8
Replying to @mmastrac
One can do with other autoregressive LLMs, see github.com/TypeLLM/TypeLLM Any feedbacks will be appreciated.
1
3
373
Introducing TypeLLM — LLMs with type-safe generation Models can still generate freely, while producing guaranteed typed outputs when structure matters. Define the output with JSON Schema. Get back values your software can use directly. Open source: github.com/TypeLLM/TypeLLM
1
1,123
Replying to @LukeberryPi
Checkout typeLLM: github.com/TypeLLM/TypeLLM
5
Replying to @CompleteSkeptic @wholyv
Or can checkout github.com/TypeAR-AI/TypeAR Type-safe generation with int, float and thinking supported.
35
Replying to @LukeberryPi
TypeAR is type-safe, but supports integer and float. Open-source: github.com/TypeAR-AI/TypeAR
5
Replying to @sydneyrunkle
Typesafe generation can go beyond classification — it can support integers and floats. Open-source: github.com/TypeAR-AI/TypeAR
188
Load more