Clover-35B-A3B-Preview vs Flint

Official Clover agentic-benchmark figure linked from the Diffuspace/Clover-35B-A3B-Preview card — DiffuSpace launch figure (card-linked agentic-benchmarks.jpg). Source · License: Apache-2.0.

Official Flint BF16 vs NVFP4 benchmark figure from the RecursiveSI/flint model card — RecursiveSI/flint card assets/flint_benchmarks_nvfp4.png. Source · License: Apache-2.0.

Verdict. Pick Clover-35B-A3B-Preview if you want DiffuSpace’s Apache-2.0 text-only diffusion-language MoE (~35B / ~3B active, 128K) derived from Qwen3.6-35B-A3B with a published Transformers + official SGLang path. Pick Flint if you want Recursive’s from-scratch Apache-2.0 39.8B / ~2.5B coding-agent MoE with INT4 / GGUF siblings and an Ollama library tag. Hardware lines stay vendor. We have no sourced bake-off. There is no Qwen3.6-35B-A3B page on this site.

Compare Clover-35B-A3B-Preview and Flint.

Dimension	Clover-35B-A3B-Preview	Flint	Source
License	Apache-2.0 (LICENSE file + card tag)	Apache-2.0 (LICENSE file + card tag)	Hugging Face LICENSE files (2026-10-31 / 2026-10-23)
Gating	No	No	Hugging Face API
Developer	DiffuSpace	Recursive	Official cards
Parameters	35B total / ~3B active (card); HF safetensors 34,661,350,848	39.8B total / ~2.5B active (card); HF safetensors ~39.8B	Official cards / HF API
Architecture	clover_moe; derived from Qwen3.6-35B-A3B; block diffusion	From-scratch Flint MoE (model_type flint)	Official cards
Modalities	text-generation (text only)	text-generation	Hugging Face API / Clover chat template
Context length	128K (card); 131,072 (config)	262,144 (card); blog also names 512k with a 256k recommendation	Official cards
Serving named	Transformers 5.19 + custom code; official SGLang recommended	Bundled SGLang kit (B300/H200); llama-server GGUF; Ollama flint	Official cards / Ollama library
Local-quant path	BF16 only on the lead card we opened	INT4 MLX + INT4/NVFP4 GGUF + Ollama	Official cards
Hardware (vendor)	BF16 “approximately 69 GB” plus state. usedStorage ~64.6 GiB. Untested here.	1× B300 or H200 (BF16). INT4 GGUF asks 48 GB unified memory. Untested here.	Official cards; HF usedStorage
Hugging Face likes	176 on the BF16 card (2026-10-31)	533 on the BF16 card (2026-10-23)	Hugging Face API
Quality vs ChatGPT	Not independently tested	Not independently tested	n/a

When to choose which

Published laptop-class GGUF / Ollama path: Flint — INT4 GGUF + ollama.com/library/flint. Size labels are vendor. Not a quality rank.
Diffusion-language / block-decoding card: Clover-35B-A3B-Preview — Card documents LowConfidence block diffusion (block_size 32). Flint is a standard coding-agent decoder MoE.
Derived-from-Qwen3.6 identification: Clover-35B-A3B-Preview — Card YAML base_model Qwen/Qwen3.6-35B-A3B. Flint is from scratch (model_type flint).
Local coding-agent card with Ollama launchers: Flint — Recursive + Ollama position Flint for coding agents. Clover’s card is general text / agentic-blog, not an Ollama tag.

Read the entity pages: Clover-35B-A3B-Preview and Flint. Both hang off ChatGPT alternatives and closed LLM APIs. Flint also hangs off Claude Code alternatives and Cursor alternatives.

Why this comparison exists

Searchers who want a ~35–40B open MoE now have two ungated Apache-2.0 cards on this site: DiffuSpace’s 35B / ~3B diffusion Clover (created 2026-10-24) and Recursive’s 39.8B / ~2.5B Flint (created 2026-10-21). There is no Qwen3.6-35B-A3B page to pair with Clover. This is identification, not a bake-off.

Sourced snapshot (2026-10-31 / Flint 2026-10-23)

Field	Clover-35B-A3B-Preview	Flint
Host	huggingface.co/Diffuspace/Clover-35B-A3B-Preview (not gated)	huggingface.co/RecursiveSI/flint (not gated)
HF likes / downloads	176 / 1,505 (2026-10-31)	533 / 420 (2026-10-23)
License	Apache-2.0 (LICENSE file)	Apache-2.0 (LICENSE file)
Size	35B / ~3B active (card); ~34.66B safetensors	39.8B / ~2.5B active
Architecture	clover_moe; Qwen3.6-35B-A3B base; block diffusion	flint (from scratch)
Job	text-generation	text-generation; coding agents
Context	128K (card)	262,144 (card)
Official serve path	Transformers + official SGLang repo	SGLang kit; GGUF; Ollama
Local-quant path	BF16 on the lead card	INT4 MLX + INT4/NVFP4 GGUF
Quality vs ChatGPT	Not independently tested	Not independently tested

License

Both Hugging Face LICENSE files are Apache License 2.0. Clover’s appendix names Copyright 2026 Alibaba Cloud (upstream Qwen). Flint’s card also tags apache-2.0 on INT4 / GGUF siblings. Not legal advice.

How to choose

Need Apache-2.0 diffusion 35B-A3B with a Qwen3.6 base: Clover-35B-A3B-Preview.
Need Apache-2.0 40B-A2.5B with GGUF / Ollama: Flint.
Need the MIT 122B coding card: Ornith-1.5-122B — Flint vs Ornith-1.5-122B.
Need Liquid’s ~2B omni (speech out): LFM3-Omni-2B.

See also AI Chat & Agents and Nemotron 3.5 Bolt (local-agent VLM, different job).

FAQ

Which is easier to host on a workstation?

Flint publishes INT4/GGUF siblings and an Ollama tag. Clover’s lead card is BF16 at about 69 GB (card) with no GGUF sibling we opened. We have not measured either VRAM.

Are both Apache-2.0?

Yes on the Hugging Face LICENSE files we opened. Clover’s appendix copyright line is “Copyright 2026 Alibaba Cloud” (upstream Qwen grant). Flint’s LICENSE is Apache 2.0 on RecursiveSI/flint. Not legal advice.

Which is better at coding or agent benches?

Flint prints self-reported Terminal-Bench / SWE-bench rows. Clover’s card points at a blog figure and does not print numeric scores we can copy. We do not merge those into a ranking.

Why isn’t this Clover vs Qwen3.6-35B-A3B?

This site has no Qwen3.6-35B-A3B page. Flint is the same-site ~40B-class Apache-2.0 MoE we can source. See also Flint vs Ornith-1.5-122B.

Sources & verification

Last verified 2026-10-31 · Last updated 2026-10-31· Edited by forkalt editors · Editorial policy

License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.