Attention ISN'T all you need?! New Qwen3 variant Brumby-14B-Base leverages Power Retention technique

When the transformer structure was launched in 2017 within the now seminal Google paper “Attention Is All You Need,” it turned an prompt cornerstone of contemporary synthetic intelligence.

Each main massive language mannequin (LLM) — from OpenAI’s GPT collection to Anthropic’s Claude, Google’s Gemini, and Meta’s Llama — has been constructed on some variation of its central mechanism: consideration, the mathematical operation that permits a mannequin to look again throughout its total enter and resolve what data issues most.

Eight years later, the identical mechanism that outlined AI’s golden age is now exhibiting its limits. Consideration is highly effective, however additionally it is costly — its computational and reminiscence prices scale quadratically with context size, creating an more and more unsustainable bottleneck for each analysis and business. As fashions purpose to motive throughout paperwork, codebases, or video streams lasting hours or days, consideration turns into the structure’s Achilles’ heel.

On October 28, 2025, the little-known AI startup Manifest AI introduced a radical alternative. Their new mannequin, Brumby-14B-Base, is a retrained variant of Qwen3-14B-Base, one of many main open-source transformer fashions.

However whereas many variants of Qwen have been educated already, Brumby-14B-Base is novel in that it abandons consideration altogether.

As an alternative, Brumby replaces these layers with a novel mechanism referred to as Energy Retention—a recurrent, hardware-efficient structure that shops and updates data over arbitrarily lengthy contexts with out the exponential reminiscence development of consideration.

Skilled at a said value of simply $4,000, the 14-billion-parameter Brumby mannequin performs on par with established transformer fashions like Qwen3-14B and GLM-4.5-Air, attaining near-state-of-the-art accuracy on a variety of reasoning and comprehension benchmarks.

From Consideration to Retention: The Architectural Shift

The core of Manifest AI’s innovation lies in what they name the Energy Retention layer.

In a conventional transformer, each token computes a set of queries (Q), keys (Ok), and values (V), then performs a matrix operation that measures the similarity between each token and each different token—primarily a full pairwise comparability throughout the sequence.

That is what offers consideration its flexibility, but additionally what makes it so expensive: processing a sequence twice as lengthy takes roughly 4 occasions the compute and reminiscence.

Energy Retention retains the identical inputs (Q, Ok, V), however replaces the worldwide similarity operation with a recurrent state replace.

Every layer maintains a reminiscence matrix S, which is up to date at every time step in response to the incoming key, worth, and a discovered gating sign.

The method seems to be extra like an RNN (Recurrent Neural Community) than a transformer: as an alternative of recomputing consideration over your complete context, the mannequin constantly compresses previous data right into a fixed-size latent state.

This implies the computational value of Energy Retention doesn’t develop with context size. Whether or not the mannequin is processing 1,000 or 1,000,000 tokens, the per-token value stays fixed.

That property alone—constant-time per-token computation—marks a profound departure from transformer conduct.

On the identical time, Energy Retention preserves the expressive energy that made consideration profitable. As a result of the recurrence entails tensor powers of the enter (therefore the title “energy retention”), it could possibly signify higher-order dependencies between previous and current tokens.

The result’s an structure that may theoretically retain long-term dependencies indefinitely, whereas remaining as environment friendly as an RNN and as expressive as a transformer.

Retraining, Not Rebuilding

Maybe essentially the most placing side of Brumby-14B’s coaching course of is its effectivity. Manifest AI educated the mannequin for under 60 hours on 32 Nvidia H100 GPUs, at a value of roughly $4,000 — lower than 2% of what a standard mannequin of this scale would value to coach from scratch.

Nevertheless, because it relied on a transformer-based mannequin, it is secure to say that this advance alone won’t finish the transformer AI-era.

As Jacob Buckman, founding father of Manifest AI, clarified in an electronic mail to VentureBeat: “The power to coach for $4,000 is certainly solely attainable when leveraging an present transformer mannequin,” he stated. “Brumby couldn’t be educated from scratch for that worth.”

Nonetheless, Buckman emphasised the importance of that outcome: “The rationale that is necessary is that the power to construct on the weights of the earlier era of mannequin architectures is a important accelerant for the adoption of a brand new modeling paradigm.”

He argues this demonstrates how attention-free programs can catch as much as transformer efficiency “for orders-of-magnitude much less” funding.

Within the loss curves launched by Manifest AI, Brumby’s coaching loss shortly converges to that of the Qwen3 baseline inside 3,000 coaching steps, even because the structure diverges considerably from its transformer origins.

Though Brumby-14B-Base started life as Qwen3-14B-Base, it didn’t stay equivalent for lengthy. Manifest AI basically altered Qwen3’s structure by eradicating its consideration layers—the mathematical engine that defines how a transformer mannequin processes data—and changing them with their new “energy retention” mechanism. This modification restructured the mannequin’s inside wiring, successfully giving it a brand new mind whereas preserving a lot of its prior data.

Due to that architectural swap, the prevailing Qwen3 weights not match completely. They had been educated to function inside a transformer’s consideration dynamics, not the brand new retention-based system. Consequently, the Brumby mannequin initially “forgot” methods to apply a few of its discovered data successfully. The retraining course of—about 3,000 steps of extra studying—served to recalibrate these weights, aligning them with the facility retention framework with out having to start out from zero.

A useful method to consider that is to think about taking a world-class pianist and handing them a guitar. They already perceive rhythm, concord, and melody, however their palms should be taught solely new patterns to provide the identical music. Equally, Brumby needed to relearn methods to use its present data by means of a brand new computational instrument. These 3,000 coaching steps had been, in impact, its crash course in guitar classes.

By the tip of this quick retraining part, Brumby had regained its full efficiency, reaching the identical accuracy as the unique Qwen3 mannequin. That fast restoration is what makes the outcome so important: it exhibits that an attention-free system can inherit and adapt the capabilities of a transformer mannequin with solely a fraction of the coaching time and price.

The benchmark development plots present an analogous pattern: the mannequin quickly approaches its goal accuracy on core evaluations like GSM8K, HellaSwag, and MMLU after only some thousand steps, matching and even barely surpassing Qwen3 on a number of duties.

Benchmarking the Brumby

Throughout normal analysis duties, Brumby-14B-Base constantly performs at or close to parity with transformer baselines of comparable scale.

Whereas it lags barely behind transformers on knowledge-heavy evaluations like MMLU-Professional, it matches or outperforms them on mathematical reasoning and long-context reasoning duties—exactly the place consideration architectures are inclined to falter. This sample reinforces the concept that recurrent or retention-based programs might maintain a structural benefit for reasoning over prolonged temporal or logical dependencies.

{Hardware} Effectivity and Inference Efficiency

Brumby’s energy retention design gives one other main benefit: {hardware} effectivity.

As a result of the state replace entails solely native matrix operations, inference could be carried out with linear complexity in sequence size.

Manifest AI reviews that their quickest kernels, developed by means of their in-house CUDA framework Vidrial, can ship hundreds-fold speedups over consideration on very lengthy contexts.

Buckman stated the alpha-stage Energy Retention kernels “obtain typical {hardware} utilization of 80–85%, which is larger than FlashAttention2’s 70–75% or Mamba’s 50–60%.”

(Mamba is one other rising “post-transformer” structure developed by Carnegie Mellon scientists again in 2023 that, like Energy Retention, seeks to get rid of the computational bottleneck of consideration. It replaces consideration with a state-space mechanism that processes sequences linearly — updating an inside state over time moderately than evaluating each token to each different one. This makes it much more environment friendly for lengthy inputs, although it usually achieves decrease {hardware} utilization than Energy Retention in early assessments.)

Each Energy Retention and Mamba, he added, “expend meaningfully fewer whole FLOPs than FlashAttention2 on lengthy contexts, in addition to far much less reminiscence.”

In response to Buckman, the reported 100× speedup comes from this mixed enchancment in utilization and computational effectivity, although he famous that “we have now not but stress-tested it on production-scale workloads.”

Coaching and Scaling Economics

Maybe no statistic within the Brumby launch generated extra consideration than the coaching value.

A 14-billion-parameter mannequin, educated for $4,000, represents a two-order-of-magnitude discount in the price of basis mannequin growth.

Buckman confirmed that the low value displays a broader scaling sample. “Removed from diminishing returns, we have now discovered that ease of retraining improves with scale,” he stated. “The variety of steps required to efficiently retrain a mannequin decreases with its parameter depend.”

Manifest has not but validated the price of retraining fashions at 700B parameters, however Buckman projected a variety of $10,000–$20,000 for fashions of that magnitude—nonetheless far under transformer coaching budgets.

He additionally reiterated that this method might democratize large-scale experimentation by permitting smaller analysis teams or firms to retrain or repurpose present transformer checkpoints with out prohibitive compute prices.

Integration and Deployment

In response to Buckman, changing an present transformer right into a Energy Retention mannequin is designed to be easy.

“It’s easy for any firm that’s already retraining, post-training, or fine-tuning open-source fashions,” he stated. “Merely pip set up retention, change one line of your structure code, and resume coaching the place you left off.”

He added that after solely a small variety of GPU-hours, the mannequin usually recovers its unique efficiency—at which level it positive aspects the effectivity advantages of the attention-free design.

“The ensuing structure will allow far sooner long-context coaching and inference than beforehand,” Buckman famous.

On infrastructure, Buckman stated the primary Brumby kernels are written in Triton, suitable with each NVIDIA and AMD accelerators. Specialised CUDA kernels are additionally out there by means of the group’s in-house Vidrial framework. Integration with vLLM and different inference engines stays a piece in progress: “We have now not but built-in Energy Retention into inference engines, however doing so is a serious ongoing initiative at Manifest.”

As for distributed inference, Buckman dismissed considerations about instability: “We have now not discovered this issue to be exacerbated in any method by our recurrent-state structure. In reality, context-parallel coaching and GPU partitioning for multi-user inference each turn into considerably cleaner technically when utilizing our method.”

Mission and Lengthy-Time period Imaginative and prescient

Past the engineering particulars, Buckman additionally described Manifest’s broader mission. “Our mission is to coach a neural community to mannequin all human output,” he stated.

The group’s objective, he defined, is to maneuver past modeling “artifacts of intelligence” towards modeling “the clever processes that generated them.” This shift, he argued, requires “basically rethinking” how fashions are designed and educated—work that Energy Retention represents solely the start of.

The Brumby-14B launch, he stated, is “one step ahead in an extended march” towards architectures that may mannequin thought processes constantly and effectively.

Public Debate and Business Reception

The launch of Brumby-14B sparked fast dialogue on X (previously Twitter), the place researchers debated the framing of Manifest AI’s announcement.

Some, together with Meta researcher Ariel (@redtachyon), argued that the “$4,000 basis mannequin” tagline was deceptive, because the coaching concerned reusing pretrained transformer weights moderately than coaching from scratch.

“They shuffled across the weights of Qwen, fine-tuned it a bit, and referred to as it ‘coaching a basis mannequin for $4k,’” Ariel wrote.

Buckman responded publicly, clarifying that the preliminary tweet had been a part of an extended thread explaining the retraining method. “It’s not like I used to be being misleading about it,” he wrote. “I broke it up into separate tweets, and now everyone seems to be mad concerning the first one.”

In a follow-up electronic mail, Buckman took a measured view of the controversy. “The top of the transformer period just isn’t but right here,” he reiterated, “however the march has begun.”

He additionally acknowledged that the $4,000 declare, although technically correct in context, had drawn consideration exactly as a result of it challenged expectations about what it prices to experiment at frontier scale.

Conclusion: A Crack within the Transformer’s Wall?

The discharge of Brumby-14B-Base is greater than an engineering milestone; it’s a proof of idea that the transformer’s dominance might lastly face credible competitors.

By changing consideration with energy retention, Manifest AI has demonstrated that efficiency parity with state-of-the-art transformers is feasible at a fraction of the computational value—and that the long-context bottleneck could be damaged with out unique {hardware}.

The broader implications are twofold. First, the economics of coaching and serving massive fashions might shift dramatically, decreasing the barrier to entry for open analysis and smaller organizations.

Second, the architectural variety of AI fashions might broaden once more, reigniting theoretical and empirical exploration after half a decade of transformer monoculture.

As Buckman put it: “The top of the transformer period just isn’t but right here. Our launch is only one step ahead in an extended march towards the long run.”

Source link