Agnes AI released Agnes 2.5 Pro Theta on 2026-11-15, a multimodal mixture-of-experts model with about 187.8B total parameters and 512 routed experts per MoE block. It accepts text, images and video frames, with a configured context limit of 262,144 tokens. The BF16 weights are open under Apache 2.0 and total roughly 350 GiB, and the model card names four H200-class GPUs as a practical starting point. Publisher-reported scores include 92.9 on GPQA Diamond and 90.5 on LiveCodeBench v6.
A situational 187.76B-parameter MoE language model from Agnes AI. Sits below the open-model average on our benchmarks; pick it for specific deployment constraints rather than peak quality. Newly released, so production-readiness is still being shaken out.
Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.
Access model weights, configuration files, and documentation.
Benchmark data sourced from: Hugging Face, Arena.ai, Artificial Analysis.
Average benchmark score against active parameters for every text model we track. Models higher up deliver more quality for their size.
See how different quantization levels affect VRAM requirements and quality for this model.
| Format | VRAM Required | Quality | |
|---|---|---|---|
| Q2_K | 463.7 GB | Low | |
| Q4_K_MRecommended | 503.2 GB | Good | |
| Q5_K_M | 521.9 GB | Very Good | |
| Q6_K | 544.5 GB | Excellent | |
| Q8_0 | 591.4 GB | Near Perfect | |
| FP16 | 769.8 GB | Full |
The top devices for this model at 4-bit, ranked by fit and speed.
| Device | Grade | Speed | VRAM |
|---|---|---|---|
| GIGABYTE NVIDIA Vera Rubin NVL72 GIGAPOD AI DLC RackGigabyte | SS | 2239.8 tok/s | 503.2 GB |
| ASUS ExpertCenter Pro ET900N G3ASUS | AA | 11.4 tok/s | 503.2 GB |
| Dell Pro Max with GB300Dell | AA | 11.4 tok/s | 503.2 GB |
| Gigabyte W775-V10-L01Gigabyte | AA | 11.4 tok/s | 503.2 GB |
| HP ZGX Fury AI StationHP | AA | 11.4 tok/s | 503.2 GB |
Energy cost on Intel GPU D270 (~2.4 tok/s, Q4_K_M) vs flagship API pricing.
| Source | Cost per 1M tokens |
|---|---|
Local (energy only)Agnes 2.5 Pro Theta on Intel GPU D270 · ~2.4 tok/s · 350W | $4.86 |
GPT-6.2 SolOpenAI · in $2.00 · out $10.00 | $4.40 |
Claude Fable 5.2Anthropic · in $8.00 · out $40.00 | $17.60 |
Gemini 4 NeonGoogle · in $3.00 · out $15.00 | $6.60 |
Grok 4.5xAI · in $2.00 · out $6.00 | $3.20 |
API prices blended at 70% input / 30% output.
Hardware amortisation not included. Run the full ROI calculator for payback math.
Cheapest current cloud rentals with at least 503 GB VRAM, refreshed hourly.
No current rental listing covers this model’s VRAM requirement on the providers we track.
Agnes 2.5 Pro Theta is an open-weights, multimodal mixture-of-experts (MoE) foundation model developed by Agnes AI. Released in November 2026, the model features 187.76 billion total parameters configured across 52 text decoder layers, paired with an extensive native context window of 262,144 tokens. It handles text, image, and video frame inputs under a fully permissive Apache 2.0 license, establishing itself as a premier open alternative to proprietary frontier models for local deployment and private enterprise infrastructure.
Operating as a top-tier local AI model with 187.76B parameters in 2026, Agnes 2.5 Pro Theta targets demanding developer tasks including automated software engineering, multi-turn conversational reasoning, visual document extraction, and scientific problem-solving. Publisher evaluations place the checkpoint at 92.9% on GPQA Diamond, 90.5% on LiveCodeBench v6, and 64.1% on SWE-bench Pro, positioning it as a direct competitor to high-end sparse models such as Nemotron 3.5 Super VL 120B and DeepSeek-V4.1-Flash.
Because the unquantized BF16 weights span approximately 350 GiB across 105 Safetensors shards, running the model on local hardware demands careful planning around memory bandwidth and VRAM allocation. For engineering teams evaluating self-hosted deployments, Agnes 2.5 Pro Theta pairs high-capacity expert routing with practical serving flexibility.
The core architecture of Agnes 2.5 Pro Theta is built around a fine-grained sparse MoE framework. Each of its 52 decoder blocks contains 512 routed experts alongside a dedicated shared expert branch. During token generation, the routing mechanism routes each token to the top 10 experts in addition to the shared branch.
Understanding Agnes 2.5 Pro Theta MoE efficiency requires separating memory footprint from active compute:
The input pathway ingests raw text, interleaved image resolutions, and sequential video frames through bundled vision processors before projecting embeddings into the 52-layer decoder. Sequence lengths are configurable up to 262,144 tokens, allowing the model to parse entire codebases, multi-hour video streams, or large stacks of PDF documentation without chunking or external retrieval steps. The reference implementation requires custom Transformers integration (trust_remote_code=True), PyTorch 2.14+, and CUDA 13.0 or higher.
Agnes 2.5 Pro Theta demonstrates state-of-the-art capability in logical deduction, programming, and visual comprehension.
Practitioners using Agnes 2.5 Pro Theta for coding benefit from its 90.5% score on LiveCodeBench v6 and 64.1% on SWE-bench Pro. It functions effectively as a self-hosted agent for repository-level debugging, unit test synthesis, refactoring, and translating legacy stacks into modern frameworks. Its long context window allows developers to pass hundreds of thousands of tokens of source files and dependency trees directly into the prompt.
The model achieved 92.9% on the GPQA Diamond benchmark and 56.0% on SciCode subproblem evaluations. It excels at formal scientific derivations, symbolic mathematics, algorithm design, and academic document synthesis. It can formulate multi-step hypotheses and solve domain-specific problems across physics, chemistry, and computational biology.
The vision encoder accepts static images and sequential video frames. Real-world applications include parsing complex financial statements, schematics, circuit diagrams, and UI wireframes. It handles chart-to-code generation, visual question answering, and anomaly detection over video streams.
Scoring 82.6% on IFBench and 84.7% on AA-LCR (v1.1), the model follows complex formatting constraints, strict JSON schemas, and structured tool-calling configurations. Native training emphasizes English and Chinese, with broad coverage across additional European and Asian languages.
To run Agnes 2.5 Pro Theta locally, teams must accommodate the substantial memory requirements dictated by its 187.76B parameter footprint.
The raw BF16 weights consume 349.7 GiB of memory before allocating space for activations and the key-value (KV) cache. Quantization is necessary for all environments outside of multi-GPU enterprise server racks.
| Format | Weight VRAM | Recommended Pool (Weights + 32k KV Cache) | Target Hardware Setup |
|---|---|---|---|
| BF16 / FP16 | ~350 GiB | ~420 GiB - 480 GiB | 4 × NVIDIA H200 (141GB) or 8 × A100/H100 (80GB) |
| FP8 / Q8_0 | ~190 GiB | ~240 GiB - 260 GiB | 4 × RTX 6000 Ada (48GB) or 4 × A100 (80GB) |
| Q4_K_M | ~110 GiB | ~135 GiB - 160 GiB | 2 × A100/H100 (80GB), or Apple Silicon 192GB Unified RAM |
| IQ3_M / Q3_K_S | ~85 GiB | ~105 GiB - 120 GiB | Mac Studio 128GB Unified RAM or 4 × RTX 4090 with system RAM offload |
When evaluating Agnes 2.5 Pro Theta hardware requirements, off-the-shelf single consumer GPUs cannot run this model.
For most self-hosted environments, Q4_K_M strikes the optimal balance between perplexity retention and hardware footprint. It compresses the model payload down to roughly 110 GiB while maintaining near-baseline reasoning and code generation scores. If extreme context lengths (128k to 262k tokens) are required, pair Q4_K_M weights with FP8 or Q4 KV-cache quantization to avoid out-of-memory errors during long-sequence generation.
Observed Agnes 2.5 Pro Theta tokens per second figures vary widely based on memory bus bandwidth:
To deploy the model:
ollama run agnes-2.5-pro-theta:q4_k_m) offers the quickest path for local CLI and API interactions.trust_remote_code=True for direct PyTorch evaluation. Specify device_map="auto" along with experts_implementation="grouped_mm" to properly schedule MoE layers across your available GPU partitions.Understanding where Agnes 2.5 Pro Theta fits requires comparing it with similarly scaled frontier open-weight models.
NVIDIA's Nemotron 3.5 Super VL 120B features a smaller total footprint, requiring less overall VRAM to load in FP8 or Q4 modes. However, Agnes 2.5 Pro Theta holds a decisive performance lead in software development and advanced scientific benchmarks. On LiveCodeBench v6, Agnes achieves 90.5% versus Nemotron's 88.2%, and on SWE-bench Pro, Agnes leads 64.1% over 51.8%. For pure agentic coding and mathematical proofs, Agnes justifies its higher VRAM footprint.
DeepSeek-V4.1-Flash remains a top benchmark competitor in pure scientific coding, marginally edging out Agnes on SciCode (57.6% vs. 56.0%) and high-difficulty reasoning on HLE-Diamond (25.3% vs. 16.8%). In contrast, Agnes 2.5 Pro Theta shows superior stability on general instruction compliance (82.6% on IFBench vs. 76.3%) and broad long-context analysis (84.7% on AA-LCR vs. 80.4%). Furthermore, Agnes provides native video frame processors directly in its model architecture, whereas many competing sparse checkpoints require separate external vision adapters.

Explore the Provider
Aggregate stats, leaderboard, release timeline, and benchmark coverage across every Agnes AI model we track.

Check which of your devices can run this model and how fast.

Compare hosted per-token prices before you commit to local hardware.

Break-even math for self-hosting, renting a GPU, or paying per token.
Live status for the hosted APIs you might use instead of running this locally.