Flint vs Ornith-1.5-122B

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.

Official Ornith-1.5-122B benchmark figure from the Hugging Face model card — Ornith-1.5-122B card assets/ornith_122b_eval.png. Source · License: MIT.

Verdict. Pick Flint if you want Recursive’s Apache-2.0 39.8B/2.5B from-scratch coding MoE with published INT4/GGUF siblings and an Ollama library tag. Pick Ornith-1.5-122B if you want a MIT 122B MoE the card says serves on one 80 GB GPU, with official vLLM/SGLang Qwen3 tool-call recipes. Hardware lines and all bench rows stay vendor. We have no sourced bake-off. Qwen3.5-35B-A3B is not a page on this site.

Compare Flint and Ornith-1.5-122B.

Dimension	Flint	Ornith-1.5-122B	Source
License	Apache-2.0 (LICENSE file + card tag)	MIT (HF card tag; LICENSE blob 404 on raw fetch)	Hugging Face cards (2026-10-23 / 2026-10-22)
Gating	No	No	Hugging Face API
Developer	Recursive	Ornith	Official cards
Parameters	39.8B total / ~2.5B active (card); HF safetensors ~39.8B	122B MoE on the card; HF safetensors ~125.1B	Official cards / HF API
Architecture	From-scratch Flint MoE (model_type flint)	qwen3_5_moe (HF tags); card says Ornith-1.0 built on Qwen3.5 / Gemma4	Official cards
Modalities	text-generation	text-generation	Hugging Face API
Context length	262,144 (card); blog also names 512k with a 256k recommendation	262,144 native; YaRN factor 4.0 ~1M (vendor)	Official cards / Recursive blog
Serving named	Bundled SGLang kit (B300/H200); llama-server GGUF; Ollama flint	vLLM ≥ 0.19.1; SGLang ≥ 0.5.9; Transformers ≥ 5.8.1	Official cards / Ollama library
Hardware (vendor)	1× B300 or H200 (BF16). INT4 GGUF asks 48 GB unified memory. Untested here.	Single 80 GB GPU (card). BF16 usedStorage ~233 GiB. Untested here.	Official cards; HF usedStorage
Hugging Face likes	533 on the BF16 card (2026-10-23)	129 on the BF16 card (2026-10-22)	Hugging Face API
Quality vs Cursor / Claude Code	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.
Published vLLM one-liner with Qwen3 parsers: Ornith-1.5-122B — Ornith’s card publishes vLLM/SGLang flags. Flint’s BF16 card ships an SGLang kit and says Transformers code is not included yet.
Smaller active-parameter MoE (card): Flint — ~2.5B active vs Ornith’s 122B card with no published active count.
Larger 122B MIT coding/agent card: Ornith-1.5-122B — Identification of size and license, not a SWE-bench ranking.

Read the entity pages: Flint and Ornith-1.5-122B. Both hang off ChatGPT alternatives and closed LLM APIs. Flint also hangs off Claude Code alternatives and Cursor alternatives. Ornith hangs off the Claude Code hub.

Why this comparison exists

Searchers who want a local coding-agent MoE now have two ungated cards on this site: Recursive’s 39.8B/2.5B Flint (created 2026-10-21) and Ornith’s 122B 1.5 card (created 2026-10-16). There is no Qwen3.5-35B-A3B page to pair with Flint. This is identification, not a bake-off.

Sourced snapshot (2026-10-23 / Ornith 2026-10-22)

Field	Flint	Ornith-1.5-122B
Host	huggingface.co/RecursiveSI/flint (not gated)	huggingface.co/ornith-ai/Ornith-1.5-122B (not gated)
HF likes / downloads	533 / 420 (2026-10-23)	129 / 742 (2026-10-22)
License	Apache-2.0 (LICENSE file)	MIT (card tag)
Size	39.8B / ~2.5B active	122B MoE (card); ~125.1B safetensors
Architecture	flint (from scratch)	qwen3_5_moe
Job	text-generation; coding agents	text-generation; coding/agents
Context	262,144 (card)	262,144 native
Official serve path	SGLang kit; GGUF; Ollama	vLLM / SGLang recipes
Local-quant path	INT4 MLX + INT4/NVFP4 GGUF	FP8 + GGUF siblings
Quality vs Cursor	Not independently tested	Not independently tested

License

Flint’s LICENSE file is Apache License 2.0. Ornith’s card tags MIT; its YAML license_link 404’d as a raw file on 2026-10-22. Not legal advice.

How to choose

Need Apache-2.0 40B-A2.5B with GGUF / Ollama: Flint.
Need the MIT 122B card and a vLLM / SGLang Qwen3-parser one-liner: Ornith-1.5-122B. Engine page: vLLM.
Need a 35B science/agent VLM instead: Sage-A1.5 — Sage-A1.5 vs Ornith-1.5-122B.
Need Apache-2.0 US text MoE at 501B/23B: Reflection Beam.

Same-size Apache-2.0 diffusion card added 2026-10-31: Clover-35B-A3B-Preview vs Flint.

See also AI Chat & Agents and Dev & Coding.

FAQ

Which is easier to host on a workstation?

Flint publishes INT4/GGUF siblings and an Ollama tag. Ornith’s card claims a single 80 GB GPU; BF16 usedStorage is still ~233 GiB. We have not measured either VRAM.

Can I serve both with vLLM?

Ornith’s card publishes a vLLM ≥ 0.19.1 recipe. Flint’s BF16 card points at a bundled SGLang kit and says Transformers modeling code is not included yet; the Recursive blog also names vLLM. Untested here.

Which is better at SWE-bench?

Both cards print self-reported tables on different harness dates and opponent columns. We do not merge them into a ranking.

Why isn’t this Flint vs Qwen3.5-35B-A3B?

This site has no Qwen3.5-35B-A3B page. Flint is a from-scratch Flint MoE, not a Qwen3.5 fine-tune. Ornith is the same-site coding/agent card we can source.

Sources & verification

Last verified 2026-10-23 · Last updated 2026-10-23· Edited by forkalt editors · Editorial policy

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