Flint 40B (MoE) vs Nemotron 3.5 Lightning 30B-A3B

Analysis

Flint 40B and Nemotron 3.5 Lightning 30B-A3B are close cousins in intent: both are sparse mixture-of-experts models with about 3B parameters active per token, both are built for coding and tool-using agents, and both are sized so that a single 32 GB card or a 48 to 64 GB Mac is the realistic home for the Q4_K_M file. Nemotron arrived on 11 August 2026 under NVIDIA's OpenMDW-1.1 license and Flint on 21 October 2026 under Apache 2.0. They keep their memory small by different routes, and that, more than the benchmark table, decides which one fits where.

Flint routes each token to 6 of 256 experts plus a shared one, about 2.5B of its 39.8B parameters; Nemotron routes to 6 of 128 plus a shared one, about 3B of its 30B. The cache designs differ more than the routing does. Flint keeps all 49 attention layers but caps 35 of them at a 513-token window, leaving 14 that grow with context (4 KV heads of width 128). Nemotron replaces most of its stack with Mamba-2 state-space layers and MoE feed-forward blocks and keeps only 6 of its 52 layers as attention (2 KV heads of width 128). Nemotron's cache is therefore the smaller: at 262,144 tokens Flint's is 7.6 GB against 1.6 GB, and Nemotron's native window continues to 1,048,576 tokens where Flint's documented window stops at 262,144 (its launch post recommends staying within 256k). The weights narrow the gap. Flint is the bigger model, 79.6 GB at BF16 against about 66 GB, yet the published Q4_K_M files are almost the same size, 25.35 GB and 25.48 GB in bartowski's repos, and this site counts 24.2 GB for Flint and 25.5 GB for Nemotron. Add the cache and the full-window totals are 35.6 GB for Flint and 30.3 GB for Nemotron, the difference between spilling out of a 32 GB RTX 5090 and fitting it. At a short context the order reverses: this site's estimates put Flint at 194 tok/s on an RTX 5090 at Q4_K_M with 8k of context against 169.2 for Nemotron, and both fit. On the benchmarks both models report, Flint leads: SWE-bench Verified 74.2 against 52.8 and Terminal-Bench 2.1 54.9 against 23.46. Treat that as a direction, not a margin, because each number comes from its own vendor's harness (Recursive used mini-swe-agent and Terminus 2, and its card counts Terminal-Bench solutions found through web retrieval as failures), and Nemotron also reports MMLU-Pro 81.62 and GPQA Diamond 75.57, which Flint's card does not. NVIDIA positions Nemotron for agent throughput, with its own claim of up to 4x higher output than similarly sized open models, and Recursive makes a speed claim for Flint too, more than 125 tokens per second on an M4 Pro MacBook Pro at short contexts. Neither company's figure has been independently reproduced. Tooling is a wash: both have an Ollama tag (flint:40b at 23 GB, nemotron-3.5-lightning:30b at 25 GB) and an NVFP4 build for Blackwell cards (22.9 GB from Recursive, 21.6 GB from NVIDIA), and both are text only.

Bottom line: Choose Flint 40B if the job is repository-level coding and you want the stronger vendor-reported coding scores, a short to medium context, and an Apache 2.0 license; it is also the faster of the two at 8k tokens in this site's estimates. Choose Nemotron 3.5 Lightning if you need a context window past 262,144 tokens, or you routinely fill a long window on a 32 GB card: its cache is a fraction of Flint's, so it holds the full 262,144-token window in 30.3 GB where Flint needs 35.6 GB. Every benchmark here is a vendor claim and Flint was a day old when this page was written, so run both on your own repository before committing.