Model library / Kwaipilot
KAT-Coder-V2.5-Dev is a 34.7 billion-parameter MOE language model for agentic coding tasks.
Mixture of experts, count not published
Context length in tokens, as published
Weights at the released precision
Licence declared on the repository
Overview
KAT-Coder-V2.5-Dev is an open-weight mixture-of-experts model released by Kwaipilot. The publisher describes it as an improvement over KAT-Coder-V2.5, trained with supervised fine-tuning and reinforcement learning to reduce abnormal behaviours and to achieve strong results on agentic coding tasks. It uses the Qwen3_5MoeForConditionalGeneration architecture.
In practice the model contains about 35 billion total parameters with roughly 3 billion activated per inference step. It supports a context window of 262 144 tokens and is distributed under the Apache-2.0 licence. The expert count and weight format are not published, so deployment must accommodate an unspecified MOE configuration.
Specification
| Repository | Kwaipilot/KAT-Coder-V2.5-Dev |
| Publisher | Kwaipilot |
| Published | 23 July 2026 |
| Architecture | Qwen3_5MoeForConditionalGeneration |
| Model type | qwen3_5_moe |
| Parameters | 34,660,610,688 |
| Layers | 40 |
| Hidden size | 2,048 |
| Attention heads | 16, 2 key/value heads |
| Context length | 262,144 tokens |
| Vocabulary | 248,320 tokens |
| Licence | apache-2.0 |
| Base model | Qwen3.6-35B-A3B |
Memory
Parameter count multiplied by the bytes each format uses. The key-value cache comes on top.
| Precision | Weights | Notes |
|---|---|---|
| bf16 | 64.6 GiB | Full precision as released |
| fp8 | 32.3 GiB | 8-bit, near-lossless on most models |
| int4 | 16.1 GiB | 4-bit, smallest footprint |
The cache costs 80 KiB per token at 16-bit, so the full 262,144-token context of one request needs about 20.0 GiB. Concurrency multiplies that number, not the weights.
Hardware
At the released precision, with 90% of the card given to the server and 2.5 GiB kept for the runtime. The context column is what is left for the cache on that setup.
| GPU | Cards needed | Free for cache | Context it holds |
|---|---|---|---|
| RTX 4090 24GB | 4 | 11.8 GiB | about 155,183 tokens |
| RTX 5090 32GB | 4 | 40.6 GiB | the full 262,144 tokens |
| L40S 48GB | 2 | 16.8 GiB | about 220,719 tokens |
| A100 80GB | 1 | 4.9 GiB | about 64,744 tokens |
| H100 80GB | 1 | 4.9 GiB | about 64,744 tokens |
| RTX PRO 6000 96GB | 1 | 19.3 GiB | about 253,487 tokens |
| H200 141GB | 1 | 59.8 GiB | the full 262,144 tokens |
Serving
A vLLM launch line for the shape above. Check the model card for a runtime the publisher recommends.
vllm serve Kwaipilot/KAT-Coder-V2.5-Dev \ --max-model-len 151552 \ --gpu-memory-utilization 0.90
Questions
The weights take about 65 GiB at the released precision. That fits on one A100 80GB, which leaves roughly 5 GiB for the key-value cache.
The repository declares apache-2.0. Read the licence text before commercial use: the name alone does not tell you what is allowed.
The configuration allows 262,144 tokens. Whether the whole window is usable depends on the memory left for the cache, which the hardware table works out per card.
This page is a reference, and Aiqre does not serve this model today. We deploy open-weight models on dedicated EU hardware on request, so if you want an OpenAI-compatible endpoint for it, email [email protected] with the model name and roughly what volume you expect.
Also in the library
Pantheon-Reasoning-26B-A4B-1.1-V2 is a 26.5 billion-parameter Gemma-4 mixture-of-experts finetune aimed at role-play scenarios, i…
Nex-N2.5-mini is a 35.1 billion-parameter mixture-of-experts language model released under Apache-2.0.
K2-Horizon-32B-Stage1 is a 32-billion-parameter dense decoder-only model released by IFM for benchmarking agentic, coding and rea…
K2-Horizon-MoVA-36B-A4B is a sparse Mixture-of-Experts language model with MoVA attention, released by IFM.
Hosting
Tell us the model and roughly what volume you expect. We reply with a price and an OpenAI-compatible endpoint, with your prompts never stored.