New architecture request: Recursive Flint (RecursiveSI/flint — 39.8B MoE coding model, ~2.5B active) #786

Before you start

• I have checked the supported models and this checkpoint is not listed there. A listed checkpoint that fails is a bug; use a Bug report instead.
• I have read the Roadmap and this model is not already planned there.
• I have read the FAQ.
• I have searched existing issues and found no duplicate.
• This is not a GGUF checkpoint. GGUF support is on the Roadmap; please do not open an issue for it.

(The ask below is the architecture, via the BF16 safetensors checkpoint. The GGUF file is only mentioned as context.)

Hugging Face link

https://huggingface.co/RecursiveSI/flint (BF16 checkpoint, model config, tokenizer, chat template, SGLang serving kit)

Is the model architecture already supported

No, this is a new model architecture

Is the quantization already supported

Not quantized (BF16 checkpoint; an NVFP4 quantization of the same architecture also exists, see "Anything else")

Error or log

Tried the released NVFP4 GGUF of the same architecture on FreeToken Desktop v0.2.0+g425dc7bf1 (Windows, CUDA 13 build). Startup refuses the GGUF metadata before any weights load, which confirms the architecture itself has no family in 0.2.0:

ValueError: GGUF architecture 'flint' is not supported (known: ['gemma4', 'glm4moe', 'glm5-next', 'glm5next', 'llama', 'minimax-m2', 'minimax-m3', 'muse-glimmer', 'qwen2', 'qwen3', 'qwen35', 'qwen35moe', 'qwen3moe', 'qwen4exp'])

I have not tried the BF16 checkpoint (single 10 GB GPU here), so this report carries no safetensors load log.

Anything else

Flint is a coding-agent MoE from Recursive (release post): 74.2% SWE-bench Verified, 89.7% SWE-bench Pro v2. Its ~2.5 B active parameters per token look like a good fit for FreeToken's expert-offload strategies on small-VRAM hosts.

Architecture facts from the model card:

• 39.8 B total parameters, ~2.5 B active per token (including the output head)
• 49 layers: 2 dense + 47 MoE; 256 routed experts, 6 selected per token, plus 1 always-active shared expert
• Attention: 32 query heads, 4 KV heads, head dim 128; per-head QK RMSNorm and learned per-head attention sinks
• 35 sliding-window layers (window 513 tokens, RoPE) + 14 full-attention layers with no positional encoding
• Vocabulary 150,144; context 262,144

Other engines that already run it: llama.cpp (a release with Flint support serves the NVFP4 GGUF, 22.9 GB, declared architecture: flint), and the SGLang serving kit shipped in the HF repo. Once the family lands, the GGUF reader would only need to map arch=flint to it.