Nemotron 3.5 Bolt 24GB vs LFM2.5 24B-A2B: Two Ways to Fit an Agent in 24GB

Here is a problem for anyone comparing these two. N​VIDIA N​emotron 3.5 Bolt 74B-A9B 24GB and LFM2.5-24B-A2B both published benchmark tables that include the same three baseline models — Qwen3.6 35B-A3B, Gemma 4 26B-A4B, and N​emotron 3.5 Lightning 30B-A3B — and the two tables disagree about what those baselines score. On PinchBench, N​VIDIA lists Qwen3.6 at 88.07 and Lightning at 83.43. Liquid AI lists the same two models at 89.93 and 84.20. Same benchmark, same named entries, different numbers, both published this month.

That is not a scandal. It is what running a harness yourself looks like, and it is the reason this comparison is written around what is verifiable rather than around a head-to-head score. The two models are also, in a sense, answers to the same question solved from opposite ends.

The constraint both models were built against

Both are 24-gigabyte models. Neither lab says so in exactly those words, and the number arrives differently in each case.

Bolt is 74B total with 9B active. N​VIDIA's Star Elastic architecture search was given one budget to satisfy — below 24GB at batch size 1 and 128K context — and the resulting checkpoint is what that constraint produced, with expert weights at 2.0625 bits per element. The memory limit is the design input, and the parameter count is the output.

LFM2.5-24B-A2B is 24B total with 2.45B active, and memory is a function of the file you choose rather than of the architecture. The published GGUF repository offers Q4_K_M at 14.53GB, a quantisation-aware QAD-Q4_0 at 13.58GB, Q5_K_M at 17.03GB, Q6_K at 19.69GB and Q8_0 at 25.49GB. The 24 in its name is a parameter count, not a footprint; the footprint is a menu.

That distinction decides a lot. Bolt has exactly one shipping configuration for inference, and there is nothing to quantise because the model arrived quantised. LFM2.5-24B-A2B has a native BF16 checkpoint at 47.96GB that you would not run on a laptop and a set of community and first-party builds that you would, with quality left as your judgement call.

What each one is actually made of

• Parameters — Bolt: 74B total, 9B active per token. LFM2.5-24B-A2B: 24B total, 2.45B active per token.

• Architecture — Bolt: Mamba-2 state-space layers interleaved with mixture-of-experts and attention, LatentMoE routing, plus a C-RADIO v4-H vision encoder. LFM2.5-24B-A2B: 40 layers — thirty double-gated short-convolution blocks and ten grouped-query attention blocks — with 64 experts per MoE block and top-4 routing.

• Provenance — Bolt: compressed from N​emotron 3.5 Super VL 120B-A12B by router-guided architecture search, then quantised with quantisation-aware distillation. LFM2.5-24B-A2B: trained directly, from a 33-trillion-token pre-training run through a two-phase mid-training stage and a five-stage post-training pipeline that ends in reinforcement learning inside agent harnesses.

• Inputs — Bolt: text, images and video in, text out. LFM2.5-24B-A2B: text only. This is the single hardest line between the two, and it is architectural rather than a tuning difference.

• Context — Bolt: 262,144 tokens. LFM2.5-24B-A2B: 131,072 tokens, doubled from the previous generation's 32,768 alongside a vocabulary expansion from 65,000 to 128,000 tokens.

• Languages — LFM2.5-24B-A2B names sixteen, including Arabic, Korean, Thai, Hindi and Russian. Bolt's model card names English and coding languages as primary with Spanish, French, German, Italian and Japanese also supported.

• Licence — Bolt: OpenMDW-1.1. LFM2.5-24B-A2B: the LFM Open License 1.0. Neither is Apache 2.0, and both are permissive enough for commercial work and unfamiliar enough to deserve a lawyer's ten minutes.

• Runtime support at release — both list llama.cpp, vLLM and their respective edge formats; LFM2.5-24B-A2B additionally ships MLX and ONNX builds and names SGLang.

The benchmark tables, read carefully

Bolt's vendor table hands it wins on Terminal-Bench 2.1 at 50.98 against 44.38 for Qwen3.6 35B-A3B and 37.22 for Gemma 4 26B-A4B, on BrowseComp at 66.75 against 48.74, and on GPQA Diamond at 86.43 against 83.40. It also prints the rows it loses: AA-Omniscience accuracy at 16.90 against G​emma's 22.17, and PinchBench at 87.82 against Q​wen's 88.07. N​VIDIA chose the peer group as models that fit in 24GB at four-bit — a set defined by the constraint Bolt was engineered against.

Liquid AI's table for LFM2.5-24B-A2B reports LiveClaw at 48.09 against 38.19 for Qwen3.6 35B-A3B, AutomationBench at 27.69 against 8.72, and ToolSandbox at 80.71 against 75.02, while conceding GPQA Diamond at 74.42, HLE at 7.40 and SWE-Bench Verified at 51.00 to larger models. Its AA-Omniscience figure is negative, minus 24.90, which in that benchmark's scoring means it is penalised for confident wrong answers rather than scored on accuracy alone.

Both tables agree on the shape of the result even where the digits differ: each model is strong at agentic tool use relative to its size and weaker on knowledge-heavy recall. Neither has an Artificial Analysis entry, and neither has a hosted endpoint, so there is no third-party number to break the tie and no way to A/B them on your own traffic without downloading both.

One independent figure does exist for Bolt: an Intelligence Index of 21, fourth of 67 models in its class, on the same panel that scores the dense models it competes with. Liquid AI reports 184 tokens per second on an Apple M5 Max for LFM2.5-24B-A2B, and N​VIDIA reports 83.99 tokens per second for Bolt on an RTX Spark machine running llama.cpp with its LiLiCorr drafter attached. Those two speeds are the most useful numbers either lab published, and they are not comparable — different hardware, different formats, different drafter setups.

Which machine each one wants

The hardware stories overlap less than the 24GB framing suggests.

Bolt is built for N​VIDIA's own edge stack: RTX Spark laptops arriving 16 October, GeForce RTX PCs, DGX Spark, DGX Station and RTX PRO workstations, with the quantised checkpoint for llama.cpp on Spark and vLLM on DGX, and the BF16 checkpoint named for Ampere, Hopper and Blackwell. The drafter is a separate 1.13B download, and the reference vLLM recipe pins a nightly build, a kernel package from a vendor repository, and a served sequence count of one.

LFM2.5-24B-A2B is the more portable of the two on paper and says so: MLX builds for Apple Silicon, ONNX for cross-platform runtimes, GGUF for CPU inference, FP8 for GPU serving, and a DSpark drafter listed as coming soon. A quantised build at 13.58GB leaves more headroom on a 24GB card than Bolt's single 22.02GB file plus its 1.73GB vision projector.

That last arithmetic is the practical crux. Bolt's weights and projector consume 23.75GB of the 24GB the model is named after, leaving context and concurrency to compete for the remainder. LFM2.5-24B-A2B at Q4 or QAD-Q4 leaves roughly ten gigabytes for the same purpose. If your workload is long-horizon agent work with a large retrieved context, the smaller active and stored footprint is not a consolation prize.

Where the routing layer fits

Neither model is something we serve, and it is worth saying so plainly rather than dressing it up: the OrcaRouter catalogue does not carry Bolt or LFM2.5-24B-A2B, so there is no endpoint here for either and no price to compare. What both have in common is that they are the local half of a stack, and local halves are almost never alone — an agent that runs a 24GB model on the machine in front of it still reaches for something bigger when the task exceeds it.

That is the part we do handle. One key across 200-plus hosted models at the provider's list price with no markup added, automatic failover so a provider wobbling mid-request is retried rather than surfaced to whoever is waiting, and a routing DSL that expresses a fallback chain as configuration instead of as application code. If you are choosing between these two checkpoints, the decision is made on your own hardware and in your own harness — and the escalations, side queries and evaluations around it are the part that goes through a gateway.

How to choose without a shared scoreboard

Four questions settle it faster than any table would.

Does the agent need to see? If it reads screenshots, PDFs or video, Bolt is the only one of the two that can, and the question is over. If it works in text and tool calls, the multimodal capability is weight you are carrying for nothing.

How much context does a turn actually need? A quarter-million tokens against 131,072 is a real difference for repository-scale or document-scale work, and it is the one dimension where Bolt's compression bought something that does not show up in any benchmark column.

What is the machine? A discrete N​VIDIA GPU with 24GB and a tolerance for a pinned nightly vLLM favours Bolt. A Mac, or a mixed fleet where the same build has to run in more than one place, favours the model that ships MLX and ONNX next to its GGUF.

And whose numbers do you believe? Both labs have published tables in which their own model wins most rows. The honest position is that each is strong within a peer group it selected, that the disagreement between the two tables on identical baselines is a reminder about harness variance, and that a benchmark number from the lab that built the model is a claim about capability rather than a measurement of it.

The short answer

Bolt buys you sight, a quarter-million-token context and the tightest compression anyone has shipped at this size, and charges for it in two-and-a-half gigabytes of headroom and a mandatory N​VIDIA recipe. LFM2.5-24B-A2B buys you portability, a smaller working set and sixteen languages, and charges for it in modality and context.

Both released within four days of each other in October 2026 — LFM2.5-24B-A2B on 13 October, Bolt on 15 October — which means neither has an independent track record yet, neither has been measured against the other on shared hardware, and anyone claiming a winner is reading the vendor tables rather than a benchmark. Download both, run your own eval set, and keep the one that finishes the work. That is the only comparison here that has not already been published by one of the two labs.

That gap between what a scoreboard measures and what a deployment meets is the reason the record matters: the documented AI failures and agent incidents around models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.