Sapiens AI Ships Agnes-3.0-Flash: Open 33B Multimodal With 262K Context Under Apache 2.0 License

Singapore-based Sapiens AI released Agnes-3.0-Flash on September 11, the company's first open-weight multimodal model under the Apache 2.0 license, delivering 33 billion parameters and a 262,144-token context window in a package that fits on a single NVIDIA H200 or H100 GPU. The model's defining engineering decision is its architecture: rather than stacking 72 standard attention layers — and accumulating a key-value cache that grows with every added token — Agnes-3.0-Flash runs 54 of those layers as gated delta-rule recurrent units that hold a fixed-size state regardless of how long the input becomes. Only the remaining 18 layers generate the kind of context-length-dependent KV cache that makes extended-context inference expensive. That 3-to-1 recurrent-to-attention ratio is not a compromise inherited from another lab's design; it reflects a deliberate architectural bet that recurrent linear attention has matured enough to carry production workloads at scale.

The KV-Cache Problem Agnes-3.0-Flash Was Built to Shrink

To understand what is technically novel about Agnes-3.0-Flash, it helps to understand the cost that standard Transformer architectures impose at long context. In a conventional attention model, every layer must store key and value vectors for every token it has processed. As the sequence grows — from 8,000 tokens to 128,000 to 262,000 — that key-value cache accumulates across all layers simultaneously. For a large model, this can consume tens of gigabytes per attention layer at maximum context length, constraining how many concurrent requests a given GPU can serve and making single-GPU deployment at 262K tokens impractical without significant engineering overhead.

The alternative approach explored extensively in research over the past two years is linear recurrent attention: rather than attending over the full history, each recurrent layer maintains a compact fixed-size state matrix that is updated with each new token using a weighted erase-and-write rule. The state does not grow. The tradeoff is expressivity: pure recurrent layers are less capable at precise recall over very long documents than full attention, because they must compress the entire history into a bounded representation.

The solution that Agnes-3.0-Flash adopts — and that recent research consistently validates — is hybridization. By interspersing full-attention layers at regular intervals through a predominantly recurrent stack, the model retains exact attention capability for the fraction of the context that needs it while reducing the cache-generating layer count from 72 to 18. Agnes-3.0-Flash limits KV-cache accumulation to 25% of its layers, substantially reducing the memory footprint at long context compared to an equivalent all-attention model at the same parameter scale.

How the Delta-Rule Recurrent Layers Actually Work

The specific recurrent mechanism Agnes-3.0-Flash uses is a gated variant of the delta rule, a memory-update principle with roots in classical neural network learning theory. In the modern linear-attention context, the delta rule updates the layer's hidden state matrix by first erasing what it previously stored at the current key address, then writing new information. Mathematically, the state update for each token compresses to: new state = previous state times (identity minus beta times outer product of key with itself), plus beta times outer product of value with key — where beta is a learned per-token write strength.

The practical consequence is that the recurrent state corrects its own previous estimates rather than simply accumulating writes, which makes it more accurate for information that changes across the sequence. The Gated DeltaNet mechanism, introduced in an ICLR 2025 paper, adds a per-layer forgetting gate that allows the model to decay irrelevant history rather than holding it indefinitely — a mechanism borrowed from conventional gated recurrent networks and adapted for the linear-attention regime.

In Agnes-3.0-Flash's published architecture, the 54 recurrent layers carry 16 key heads and 48 value heads with a head dimension of 128, operating behind a causal convolution kernel and gated normalization. The 18 global-attention layers use a more conventional grouped-query setup — 24 query heads against 4 KV heads, a 6-to-1 ratio that reduces the size of the KV cache each attention layer generates — with RMS normalization applied to both queries and keys before the attention computation.

The feed-forward structure adds a further unusual element: every layer runs a standard SwiGLU projection to an intermediate size of 17,408, alongside a smaller parallel SwiGLU branch of size 2,048. These two paths execute simultaneously rather than sequentially. This parallel feed-forward design appears in some recent models as a way to increase expressivity per layer without proportionally increasing latency.

Multimodal Pipeline and Reasoning Controls

Agnes-3.0-Flash accepts text, images, and video through a bundled vision tower: 27 transformer layers processing patches of size 16 through a 2×2 spatial merge before projecting to the 5,120-dimensional model hidden size. The positional encoding system uses a three-axis rotary embedding — separate rotation dimensions for text, image height, and image width — to allow the model to represent spatial relationships in visual inputs without collapsing them into a purely sequential format.

The chat interface exposes three named reasoning levels — high, medium, and low — corresponding to different inference budgets, plus a thinking-off mode that disables the extended reasoning pass entirely for latency-sensitive applications. This graduated control over reasoning compute is increasingly standard in capable reasoning models and lets developers trade response quality for speed depending on the workload. Tool calling follows an explicit syntax using <tool_call> blocks that the model emits and that applications parse to execute functions and return results, consistent with function-calling conventions in other major model families.

The model ships with an OpenAI-compatible Chat Completions API as well as an Anthropic-compatible Messages API and a Responses API — three separate endpoint conventions in a single deployment, which reduces integration overhead for teams already using any of these formats. Full API documentation is available at the Agnes AI developer hub.

Benchmark Scores and What They Can and Cannot Establish

Agnes AI's model card presents a reference benchmark table comparing Agnes-3.0-Flash against nine other models including Qwen3.6-35B-A3B, Kimi K2.5, Muse Glimmer, and Gemini 3.5 Flash. The card explicitly qualifies the table: "Header parameter figures mix total and active counts, and harnesses and snapshot dates differ across sources, so treat cross-column comparisons as reference values rather than a controlled head-to-head evaluation." That self-caveat is worth taking seriously.

On GPQA Diamond — a set of 198 graduate-level multiple-choice questions in biology, chemistry, and physics, written by domain experts and designed to require genuine reasoning rather than web retrieval — Agnes-3.0-Flash posts a company-compiled score of 85.05. Domain experts holding or pursuing PhDs in the relevant fields score approximately 65-70% on this benchmark, so a score in the mid-80s is substantially above human-expert performance by this measure. The benchmark has, however, been largely saturated at the top of the 2026 leaderboard, with some proprietary models scoring above 92, so an 85 positions Agnes-3.0-Flash in the middle tier of capable 2026 models rather than at the frontier — and that characterization is consistent with what a 33 billion parameter open-weight model should realistically achieve.

On IFBench — an instruction-following evaluation measuring how reliably a model adheres to complex specifications — the company-compiled reference score is 74.20, landing above models like Qwen3.6-35B-A3B (64.4) but below Qwen3.8-27B (79.5) and MiniMax M3 (82.9) on the same table.

The Artificial Analysis Intelligence Index score of 36 carries a separate important qualification: the listing on the Artificial Analysis platform is explicitly marked as an estimate, with the note that "independent evaluation [is] forthcoming." That score — drawn from a composite of 10 evaluations including coding, agentic task completion, long-context retrieval, and scientific reasoning — reflects partial measurement rather than a completed independent audit. Output speed, measured by Artificial Analysis on the Agnes AI API, comes in at 237 tokens per second, well above the 144.8 token-per-second median for comparable-priced reasoning models. Time to first token, however, sits at 1.84 seconds against a peer median of 0.95 seconds — a latency gap that matters for interactive applications where response initiation speed is noticeable.

Where Agnes-3.0-Flash Sits in the Open-Weight Multimodal Landscape

The open-weight model landscape as of September 2026 offers several 30B-scale options with 262K-range context windows, each with different architectural and licensing tradeoffs.

Qwen3.6-35B-A3B, from Alibaba's Qwen team, is the closest architectural peer: it also uses a hybrid design combining Gated DeltaNet linear attention with full attention layers, and carries a 262K context window under Apache 2.0. Its key difference from Agnes-3.0-Flash is that it uses a mixture-of-experts structure, activating only around 3 billion of its 35 billion total parameters per token. That sparse activation sharply reduces inference compute relative to Agnes-3.0-Flash, which runs all 33 billion parameters on every forward pass. For throughput-sensitive deployments, Qwen3.6-35B-A3B may be more economical per token; for deployments prioritizing dense-model quality characteristics or where MoE routing overhead creates complications, Agnes-3.0-Flash is the alternative.

Qwen3.8-27B, the dense Qwen family member at a similar scale, offers a pure-attention design under Apache 2.0 but without the KV-cache reduction that comes from hybrid architecture. For teams that want a 30B-scale Apache-licensed model without the hybrid complexity, it remains a viable baseline.

Llama 4 Scout extends context dramatically further — to 10 million tokens in theory — but ships under the Llama 4 Community License, which excludes EU-based developers from accessing multimodal weights and requires a separate agreement for products exceeding 700 million monthly active users. Agnes-3.0-Flash's Apache 2.0 license carries no such restrictions: commercial use, redistribution, and fine-tuning are permitted without geographic exclusions or user thresholds.

The hybrid architecture itself is becoming a convergent design choice among capable open-weight models in 2026. Both Qwen3.6 and Agnes-3.0-Flash adopted gated delta-rule mechanisms as load-bearing architectural elements, moving a research technique that began as an ICLR 2025 paper into production-weight open models within roughly a year. That convergence is evidence that the tradeoff the architecture makes — bounded recurrent state in exchange for KV-cache containment — is clearing quality bars that purely theoretical arguments could not resolve.

Singapore, Apache 2.0, and the Third-Option Positioning

Sapiens AI, the Singapore company behind the Agnes AI product line, has built its public positioning on an explicit claim: that a capable AI model can originate outside the US-China duopoly that dominates frontier AI development. Founder and CEO Bruce Yang, an AI PhD researcher at the National University of Singapore and former Microsoft and LinkedIn engineer, raised a $10 million Series A in February 2026 led by LOOK FORWARD VCC, bringing total reported funding to approximately $20 million. By March 2026, the company was reporting that it was approaching $20 million in annual recurring revenue — fast monetization for a company at this funding level. Sapiens Technology announced its inclusion in the Forbes Asia 100 to Watch 2026 list via a press release on September 9, two days before this model release.

The company's commercial model has positioned the Agnes AI API as a single endpoint covering text, image, and video inference from one account and API key, priced at $0.05 per million input tokens and $0.15 per million output tokens — with all three items currently available at no cost during a free promotion period that has no disclosed end date. The open-weight release is the complement: teams that want full control over the model, including fine-tuning, offline deployment, and data sovereignty guarantees, can download the weights directly and host them independently.

This positioning fits neatly within Singapore's National AI Strategy 2.0, launched in December 2023, which identifies AI capability development as a strategic national priority. The National AI Council formed in February 2026 under Prime Minister Lawrence Wong, and the government committed more than SGD 1 billion to public AI research and talent development through 2030. Agnes-3.0-Flash arriving as an open-weight Apache 2.0 release is a commercial milestone, but it also represents the kind of verifiable capability benchmark that the strategy's broader ambition depends on — a Singapore-origin model that can be independently evaluated and deployed alongside models from Alibaba, Meta, or OpenAI.

Deployment Constraints and What Independent Evaluation Has Yet to Confirm

Agnes-3.0-Flash's hardware requirements are real constraints for many teams. Running the bf16 checkpoint requires approximately 66 gigabytes of disk space, a single H200 with 141 GB of HBM or an H100 with 80 GB, and 128 GB of host RAM. At tensor parallelism 2 — splitting across two GPUs — available context and concurrency increase, but the hardware bar rises accordingly. Teams without access to high-end accelerators will need to wait for quantized variants or use the API. The model also ships its own custom implementation and requires trust_remote_code=True when loaded through Hugging Face Transformers — standard practice for architectures that deviate significantly from the Transformer base class, but a flag that cautious deployments will want to audit before production use.

On inference throughput, the 237-token-per-second output speed is competitive for the price tier, but the 1.84-second time to first token — nearly twice the peer median — creates a visible latency gap for applications where users wait for the first token before seeing any output. This is particularly relevant for interactive agentic workflows where multiple tool-call cycles are chained, and where first-token latency compounds across turns.

The most significant outstanding uncertainty is the Intelligence Index score. Artificial Analysis's score of 36 is an estimate that reflects partial evaluation — the platform explicitly flags that independent full evaluation of Agnes-3.0-Flash is forthcoming. Until that evaluation completes, the GPQA Diamond, IFBench, and comparison-table figures remain company-compiled reference values assembled from different harnesses and model snapshots, not a controlled independent audit. The architectural novelty of Agnes-3.0-Flash means independent reproduction of its benchmark performance matters more than usual: researchers will want to verify that the delta-rule recurrent layers perform as described at 262K context on tasks requiring long-document retrieval, where the fixed-state compression of recurrent layers is most likely to diverge from full-attention behavior.

The next meaningful checkpoint for this model is when Artificial Analysis completes its full composite evaluation. That result will determine whether the reference benchmark scores in the company's table hold under standardized measurement conditions — and whether Agnes-3.0-Flash's hybrid architecture produces the inference-quality tradeoff that the research literature on Gated DeltaNet hybrids predicts. If it does, Sapiens AI will have shipped the most capable open-weight hybrid-recurrent model at the 30B scale under a genuinely permissive license, at a moment when the rest of the frontier has not yet converged on what that architecture is worth.