Liquid AI shipped two new open-weight models in the space of 48 hours this week, and the smaller of the pair is turning out to be the more interesting story. LFM2.5-24B-A2B, released on October 13, 2026, packs 24 billion total parameters but activates only 2 billion of them per token through a mixture-of-experts design. According to benchmark tracking site BenchLM, the model beats Alibaba’s Qwen3.6-35B-A3B on tool-use tasks while losing to it on SWE-Bench Verified, the industry’s standard measure of real-world coding ability. That split result is the clearest signal yet that the open-weight AI market is no longer a single race to the biggest parameter count. It is fragmenting into specialists.
The timing matters. LFM2.5-24B-A2B landed in the middle of what BenchLM’s October release tracker now counts as 22 confirmed model launches this month alone, pushing its leaderboard past 221 tracked models as of October 14. Meta’s own 816-billion-parameter Muse Spark Open arrived just two days later, on October 15. NVIDIA’s Nemotron 3.5 Super VL 120B-A12B landed the same day. The United Arab Emirates-based Institute for Machine Intelligence shipped six variants of its Jais 3 family on October 15 as well. For a market that spent most of 2025 obsessing over frontier-scale parameter counts, the question this week is whether efficiency, not sheer size, is about to become the more valuable currency.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Is LFM2.5-24B-A2B?
LFM2.5-24B-A2B is Liquid AI’s latest entry in its Liquid Foundation Model line, and the naming convention tells you most of what you need to know before reading a single benchmark. The “24B” refers to the model’s total parameter count, while the “A2B” denotes that only 2 billion of those parameters are active for any given token during inference. That 12-to-1 ratio between total and active parameters is unusually aggressive even by current mixture-of-experts standards, and it is the entire point of the release. Liquid AI has staked its identity on inference efficiency rather than raw scale, and LFM2.5 pushes that bet further than its predecessors did.
Per figures reported by Developers Digest, LFM2.5-24B-A2B reaches 184 tokens per second running locally on an Apple M5 Max chip, a benchmark environment increasingly used by model makers to demonstrate that frontier-adjacent capability no longer requires a data center. That framing echoes what Meta did with Muse Spark Open a day earlier, when it claimed the 816-billion-parameter model can run on a pair of Apple M5 Ultra Mac Studios using Apple’s MLX framework. Both companies are courting the same audience: developers who want serious model capability without renting cloud GPU clusters by the hour.
Inside the Mixture-of-Experts Design
Mixture-of-experts architecture has become the default approach for any lab trying to balance capability against cost. Instead of running every parameter in the network for every token, a MoE model routes each token to a small subset of specialized sub-networks, or “experts.” The rest of the network sits idle for that particular computation. The result is a model that behaves, in terms of raw knowledge and capability, closer to its total parameter count, while costing closer to its active parameter count to actually run.
Liquid AI is not alone in leaning on this architecture this week. NVIDIA’s Nemotron 3.5 Super VL 120B-A12B uses the same pattern at a different scale: 120 billion total parameters against 12 billion active. Meta’s Muse Spark Open goes further still, routing each token through 256 experts per MoE layer and selecting 8 of them, landing at 42 billion active parameters out of 816 billion total. The direction of travel across the industry is consistent: total parameter counts keep climbing on paper, while the parameters that actually fire during inference grow far more slowly. That is what makes it possible to talk, in the same week, about an 816-billion-parameter model and a 24-billion-parameter model both targeting desktop and workstation hardware.
Benchmark Results: Where LFM2.5 Wins and Loses
The most useful detail in this week’s release is not a victory lap. It is a genuine split decision. Developers Digest reports that LFM2.5-24B-A2B outperforms Qwen3.6-35B-A3B, Alibaba’s comparably sized open-weight model, on tool-use benchmarks, the tests that measure how reliably a model can call external functions, APIs, and agents to complete a task. That is directly relevant to Liquid AI’s stated focus on agentic workloads, where a model spends much of its time deciding which tool to invoke rather than generating prose.
But the same reporting shows LFM2.5 trailing Qwen3.6-35B-A3B on SWE-Bench Verified, the benchmark most widely cited for real-world software engineering performance because it tests whether a model can resolve actual GitHub issues rather than toy coding problems. That gap matters because coding and agentic work is precisely the use case Liquid AI is marketing LFM2.5 toward. A model that wins at deciding which tool to call but loses at actually writing the fix is a mixed pitch to the exact audience the release is aimed at.
Qwen3.6-35B-A3B’s own name signals the comparison point: Alibaba’s naming scheme for its Qwen MoE family denotes active parameter count the same way Liquid AI does, suggesting a model with roughly similar efficiency ambitions but a larger total and active footprint than LFM2.5. The two models sit close enough in design philosophy that the benchmark splits read less like a knockout and more like a genuine trade-off between tool orchestration and code generation depth.
| Metric | LFM2.5-24B-A2B (Liquid AI) | Qwen3.6-35B-A3B (Alibaba) |
|---|---|---|
| Release date | October 13, 2026 | Prior release (benchmark baseline) |
| Total parameters | 24 billion | 35 billion |
| Active parameters per token | 2 billion | Implied ~3 billion by naming convention |
| Local inference speed | 184 tokens/sec on Apple M5 Max | Not specified in available reporting |
| Tool-use benchmarks | Leads | Trails LFM2.5 |
| SWE-Bench Verified | Trails | Leads LFM2.5 |
| Primary design goal | Agentic efficiency, low active-parameter footprint | Balanced coding and reasoning |
The October 2026 Open-Model Release Wave
LFM2.5-24B-A2B did not launch into a quiet market. The same week produced a genuine pileup of open-weight releases, a pace BenchLM’s tracker puts at 22 confirmed launches for October 2026 alone, as of its October 15 update. Liquid AI itself shipped a second model, a decision-focused system called d1-ultra, on October 15, just two days after LFM2.5. The Institute for Machine Intelligence released six variants of its Jais 3 family, complete with weights and adapters, the same day. NVIDIA’s Nemotron 3.5 Super VL 120B-A12B, a 120-billion-parameter model combining reasoning, tool use, and multimodal understanding across text, images, documents, and video, also landed October 15, though its benchmark results were still pending as of initial reporting.
Meta’s Muse Spark Open, developed by Meta Superintelligence Labs, arrived in that same window, with the company describing it as “the most capable US open-weight model to date.” That phrase, confirmed in Meta’s own release materials and published on Meta’s developer blog, is notable for what it implies about where the open-weight race now sits: a US lab explicitly measuring itself against the broader open-weight field rather than only against closed competitors like OpenAI or Google. Hugging Face confirmed the model’s architecture details in its own release notes, describing an 816-billion-parameter natively multimodal mixture-of-experts system with a 1-million-token context window, available for inference through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, with weights distributed through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.
Separately, Google’s Gemini 4 Argon widened its own availability on October 13, opening up through the Gemini API and AI Studio after weeks of restricted access to trusted defenders under Google’s Fairwind Program. It is a proprietary model rather than an open-weight one, but its broader rollout in the same seven-day window underscores how compressed the current AI release cycle has become. Three open-weight launches, one proprietary expansion, and a second Liquid AI model, all inside one week.
Table: October 2026’s Open-Weight Launch Cluster
| Model | Developer | Release Date | Total Params | Active Params | Notable Detail |
|---|---|---|---|---|---|
| LFM2.5-24B-A2B | Liquid AI | Oct 13, 2026 | 24B | 2B | 184 tok/sec on Apple M5 Max |
| d1-ultra | Liquid AI | Oct 15, 2026 | Not disclosed | Not disclosed | Decision-focused model |
| Jais 3 (6 variants) | Institute for Machine Intelligence | Oct 15, 2026 | Varies by variant | Varies by variant | Weights and adapters released together |
| Nemotron 3.5 Super VL | NVIDIA | Oct 15, 2026 | 120B | 12B | Multimodal: text, images, documents, video |
| Muse Spark Open | Meta Superintelligence Labs | Oct 15, 2026 | 816B | 42B | 1M-token context; runs on two Mac Studios via MLX |
Why Liquid AI Is Betting on Efficiency Over Scale
Liquid AI has never tried to compete on total parameter count, and LFM2.5-24B-A2B continues that pattern rather than breaking from it. Where Meta is shipping an 816-billion-parameter system and NVIDIA a 120-billion-parameter one, Liquid AI’s entire pitch rests on the 12-to-1 ratio between LFM2.5’s total and active parameters, and on the claim that it can hit 184 tokens per second on consumer-grade Apple silicon rather than data center accelerators. That is a fundamentally different sales pitch than “biggest model wins.” It is “cheapest model that still competes wins,” aimed squarely at developers who want to self-host an agentic system without a six-figure GPU bill.
Shipping two models in 48 hours, LFM2.5 on October 13 and d1-ultra on October 15, also signals a company trying to establish release cadence as a competitive weapon in its own right. In a market where BenchLM is tracking 22 launches in a single month, simply being first, or being frequent, becomes part of the story independent of any individual benchmark win. Liquid AI does not have Meta’s compute budget or NVIDIA’s hardware relationships. What it has is speed and a narrow, defensible niche: small-footprint models that punch above their active-parameter weight on agentic tool use.
Historical Context: How We Got Here
The open-weight model market looked very different even twelve months ago. Through most of 2024 and early 2025, the open-weight conversation was dominated by a handful of players, chiefly Meta’s Llama series and a growing field of Chinese labs including Alibaba’s Qwen and DeepSeek. The idea of a US lab explicitly branding a release as the most capable US open-weight model would have been an odd thing to say in 2024, because the competitive framing at the time was almost entirely about catching up to closed, proprietary frontier models from OpenAI and Anthropic.
What changed in 2026 is the sheer density of releases and the diversification of design goals. Rather than every lab chasing one leaderboard, companies are now explicitly targeting different operating points: NVIDIA building for enterprise multimodal reasoning, Meta building for maximal open-weight capability with a desktop deployment story, the Institute for Machine Intelligence building a multilingual open family, and Liquid AI building for the smallest possible active-parameter footprint that still competes on specific agentic benchmarks. The Swiss-backed Apertus project, which debuted its own fully open 1.5 release with a 262,000-token context window earlier this month, is further evidence that national and regional open-weight projects are becoming a permanent fixture of the landscape rather than a one-off experiment.
Market Impact: What This Means for Developers and Enterprises
For developers building agentic applications, products that need a model to decide which API to call, which database to query, or which internal tool to invoke, LFM2.5-24B-A2B’s tool-use advantage over Qwen3.6-35B-A3B is the detail worth testing directly rather than taking on faith from a single benchmark table. Benchmarks like SWE-Bench Verified and generic tool-use suites are useful directional signals, but they are not guarantees of production performance on a specific company’s internal toolchain. Teams evaluating LFM2.5 for agentic deployment should run it against their own tool-calling workloads before committing, especially given the SWE-Bench Verified gap suggests the model’s coding depth may lag its orchestration skill.
For enterprises, the bigger signal is cost. A 2-billion-active-parameter model that hits meaningful tool-use benchmark numbers is dramatically cheaper to run at scale than a 42-billion-active-parameter model like Muse Spark Open, even if the larger model wins on raw capability. That tradeoff, cost per inference versus ceiling on capability, is becoming the central enterprise procurement question in late 2026, replacing the simpler “which model scores highest” calculus that dominated planning a year earlier. Expect procurement teams to start running side-by-side cost-per-task comparisons rather than leaderboard rank comparisons.
Competitive Landscape: Liquid AI vs the Field
Liquid AI’s position in this market is narrower than Meta’s or NVIDIA’s by design. It is not trying to be a general-purpose frontier model provider. Its competitive set is really the efficiency-focused tier: Alibaba’s Qwen3.6 family, and increasingly, Google’s smaller Gemini variants and any future compact release from Mistral, whose own Large 4 model this month took the opposite approach, launching at 1.05 trillion parameters. The contrast between Mistral’s trillion-parameter bet and Liquid AI’s 24-billion-parameter bet, both announced within weeks of each other, captures just how split the current strategic landscape has become.
Intel’s decision to offer day-zero hardware support for Meta’s Muse Spark Open, confirmed in Intel’s own newsroom announcement, is also a signal worth watching from Liquid AI’s perspective. Hardware vendors are increasingly picking favorites among open-weight releases to showcase their own silicon, and so far that spotlight has gone disproportionately to the largest models rather than the most efficient ones. If Liquid AI wants similar hardware-vendor attention for LFM2.5, it will likely need to lean harder on its Apple M5 Max performance numbers as a counter-narrative to the bigger-is-better showcase pattern.
Running LFM2.5 Locally: What the Deployment Story Looks Like
Liquid AI has followed the now-standard open-weight distribution pattern of publishing model weights for use with common inference frameworks. A typical local-deployment pattern for a Liquid Foundation Model release, using the Hugging Face Transformers library, looks like this:
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "LiquidAI/LFM2.5-24B-A2B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
device_map="auto",
torch_dtype="auto"
)
prompt = "Call the weather API for Boston and summarize the result."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(output[0], skip_special_tokens=True))That snippet reflects the conventional way developers load any Hugging Face-hosted mixture-of-experts model today; actual repository paths, quantization options, and recommended inference stacks for LFM2.5 should be confirmed against Liquid AI’s own documentation before production use, since specific serving configurations were not detailed in this week’s reporting.
What’s Next: Predictions for the Open-Model Race
- Benchmark leaderboards fragment further. Expect tracking sites like BenchLM to start publishing separate leaderboards for tool-use, coding, and multimodal reasoning rather than one composite ranking, since this week’s results show no single model winning across all three.
- Liquid AI ships a third model within weeks. Given the 48-hour gap between LFM2.5 and d1-ultra, a continued rapid-release cadence through November 2026 looks likely as the company tries to claim mindshare through frequency rather than individual benchmark dominance.
- Enterprise buyers start requesting cost-per-task benchmarks. As active-parameter counts diverge this sharply (2 billion versus 42 billion among models released this same week) procurement conversations will increasingly demand cost-normalized comparisons rather than raw capability scores.
- Hardware vendors pick more open-weight favorites. Following Intel’s day-zero support for Muse Spark Open, expect AMD, Qualcomm, or additional Apple silicon partnerships tied to specific open-weight releases before year-end.
- SWE-Bench Verified becomes a standard disclosure requirement. As more labs show split results between tool-use and coding benchmarks, expect SWE-Bench Verified scores to become a mandatory line item in release announcements, the way context-window size became standard in 2025.
How LFM2.5 Fits the Broader Efficiency Trend
It is worth stepping back from the individual benchmark scores to look at the shape of the trend line. A year ago, a 2-billion-active-parameter model competing on tool-use benchmarks against a 3-billion-active-parameter rival would have been a minor story. Today it sits alongside an 816-billion-parameter release from Meta and a 120-billion-parameter multimodal system from NVIDIA in the same news cycle, and arguably draws more practical interest from working developers than either. That is the real story behind LFM2.5-24B-A2B: not that it beats a rival on one benchmark, but that the market has matured enough to have a genuine, useful conversation about which 24-billion-parameter model to self-host, a conversation that simply did not exist with this much nuance twelve months ago.
For the open-source LLM buyer’s guide maintained on this site, LFM2.5-24B-A2B is a strong candidate addition specifically for teams prioritizing agentic tool-calling over raw coding depth, while Qwen3.6-35B-A3B remains the stronger pick for teams whose primary workload is software engineering automation. That is a genuinely new kind of recommendation to be able to make with confidence, since it did not require picking an overall best model, only the right model for a specific job.
Frequently Asked Questions
What is LFM2.5-24B-A2B?
LFM2.5-24B-A2B is an open-weight mixture-of-experts language model released by Liquid AI on October 13, 2026. It has 24 billion total parameters but activates only 2 billion per token, and Liquid AI reports it reaches 184 tokens per second running locally on an Apple M5 Max chip.
How does LFM2.5-24B-A2B compare to Qwen3.6-35B-A3B?
According to benchmark reporting from Developers Digest, LFM2.5-24B-A2B outperforms Qwen3.6-35B-A3B on tool-use benchmarks but underperforms it on SWE-Bench Verified, the standard test for real-world software engineering tasks.
Is LFM2.5-24B-A2B available for download?
LFM2.5-24B-A2B is distributed as an open-weight model, following Liquid AI’s established pattern of publishing its Liquid Foundation Model releases for use with standard inference frameworks such as Hugging Face Transformers.
What other AI models launched the same week as LFM2.5?
The same week saw NVIDIA release its Nemotron 3.5 Super VL 120B-A12B model, the Institute for Machine Intelligence launch six variants of Jais 3, Meta release its 816-billion-parameter Muse Spark Open, Liquid AI itself ship a second model called d1-ultra, and Google widen access to Gemini 4 Argon. BenchLM’s tracker counted 22 confirmed model releases across October 2026 as of its October 15 update.
Why does active parameter count matter more than total parameter count?
In mixture-of-experts models, only the active parameters are computed for any given token, so active parameter count is a much closer proxy for actual inference cost and speed than total parameter count. A model with a low active-to-total ratio, like LFM2.5’s 2-billion-to-24-billion split, can be significantly cheaper to run than a model with a similar total parameter count but a higher active fraction.
Does LFM2.5-24B-A2B support multimodal inputs?
Available reporting on LFM2.5-24B-A2B focuses on its text-based tool-use and coding benchmark performance. Multimodal capability details were not confirmed in current release coverage, unlike NVIDIA’s Nemotron 3.5 Super VL, which was explicitly described as combining reasoning and tool use with understanding of text, images, documents, and video.
Who is behind Liquid AI?
Liquid AI is the developer of the Liquid Foundation Model series, which LFM2.5-24B-A2B belongs to. The company shipped two models within 48 hours this week: LFM2.5 on October 13 and a decision-focused model called d1-ultra on October 15, 2026.
