Recursive Flint runs at 91 tok/s (90.6 as recorded) on an M3 Ultra, a community run reported by x.com (MLX, 4-bit, 1K tokens per request, batch 1). It needs about 20 GB of unified memory. It is Recursive's 39.8B mixture-of-experts model (2.5B active), released under the Apache 2.0 licence.
Want to run Recursive Flint faster — or step up to bigger models? Find your Mac →This is a community run: a real run, reported by its source and linked below. The index records it as given and never adjusts it.
| Metric | Value |
|---|---|
| Tokens per second | community91 tok/s |
| Provenance | community run · x.com |
| Quantization | 4-bit |
| Engine | MLX |
| Context | 1K tokens per request, batch 1 |
| Figure dated | 2026-10-21 |
| Minimum RAM | 20 GB |
| Parameters | 39.8B MoE (2.5B active) |
| Chip bandwidth | 819 GB/s |
Each row as the index holds it, with its source. “Mac RAM” is the memory of the Mac a run used, not the model’s requirement (20 GB). The headline figure follows the reference rule: the lower-median measured plain-decode run at 5 bits or fewer (a run whose quant was not stated counts only when no stated-quant run qualifies), else the estimate; a run marked “not used to anchor estimates” is listed here but never becomes the headline.
| Speed | Quant | Engine | Mac RAM | Context | Source | Note |
|---|---|---|---|---|---|---|
| community104.5 tok/s | INT4 | MLX | — | 19-token prompt, 512 generated, median of 5 | x.com | Mac Studio, RAM not stated. mlx-lm from the then-unmerged PR #1998 branch (merged 24 Oct, not yet released) with a load-time workaround for the uint8-stored weights, so it does not anchor; RecursiveSI/flint-int4. The same thread's 107 tok/s on Ollama's MLX engine had reasoning on and an unstated quant. 4.54 bpw = the 22.6 GB RecursiveSI/flint-int4 checkpoint over 39.8B parameters.; not used to anchor estimates |
| community90.6 tok/s | 4-bit | MLX | — | 1K tokens per request, batch 1 | x.com | mlx-vlm 0.7.10, 4-bit, no MTP or draft model (confirmed in reply 2113009118990496083); peak memory under 25 GB. RAM not restated on this chart (the BF16 chart in the same thread says 512 GB). The batch-8 aggregate is not plain decode. |
| community51.3 tok/s | BF16 | MLX | 512 GB | 1K, single request | x.com | mlx-vlm 0.7.10, BF16, no MTP or draft model; posted by mlx-vlm's author. 32.7 tok/s at 128K is a long-context figure, not this row. |
Ollama — install it (brew install ollama or the app from ollama.com), then pull the model by its tag, checked on ollama.com on 2026-10-24:
MLX — pick a build from Hugging Face’s MLX listings, then run it with its repo id:
Downloads: Ollama page → · model card →
The 91 tok/s figure above came from MLX at 4-bit; other runtimes and quants run at different speeds.
Prefer a GUI? LM Studio downloads GGUF and MLX builds in one click. Step by step: install Ollama on a Mac.
Recursive's Apache 2.0 coding-agent MoE (21 Oct 2026): 39.8B total, ~2.5B active, 262K context, text only, always reasons. Its card reports SWE-bench Verified 74.2 and Terminal-Bench 2.1 54.9 (its SWE-bench Pro v2 89.7 is an internal mini-swe-agent run, not comparable to SEAL's); the index's capability figure is estimated from these vendor coding benchmarks only. Stock llama.cpp b11803+, Ollama (flint:40b) and LM Studio run it, though llama.cpp's expert-scaling workaround costs 10–30% for now and issue #31000 (NVFP4 tool calls) is open. A ~18 GB 3-bit build fits a 24 GB Mac; community llama.cpp runs measured 92.4 tok/s on an M5 Pro and 79.6 on an M4 Pro (Q4_1).
Apache 2.0 licence · 39.8B MoE · 262K context · LLMCheck capability 37/50 · see its leaderboard rank →
Recursive Flint needs about 20 GB of unified memory. These current Apple Silicon Macs have the headroom to run it comfortably:
As an Amazon Associate, LLMCheck earns from qualifying purchases. These affiliate links cost you nothing extra and help keep the LLMCheck index free.
See how Recursive Flint stacks up against other models on your specific Mac hardware.
Open Compare Tool Full Leaderboard