26+ AI Labs. 108+ Models.
One Private Platform.

OpenAI's smartest model (released Apr 23, 2026).

OpenAI's most capable model and the top of the GPT-5.6 family (released Jul 2026).

OpenAI's flagship September 2026 frontier model with a 1.05M context window, MoE architecture, and SOTA reasoning across coding, formal mathematics, and long-horizon agentic automation.

Premium creative and analytical work

Creative writing, complex UI/UX coding, architectural decisions, production debugging, strategic advising, deep conceptual reasoning, emotional support

Successor to Opus 4.6 at the same price (released Apr 16, 2026).

Anthropic's most capable model (released May 28, 2026; knowledge cutoff Jan 2026).

Anthropic's flagship 1M-context frontier multimodal foundation model, engineered for deep reasoning, autonomous coding agents, and complex long-horizon system execution.

Claude Opus 5.5 (anthropic/claude-opus-5.5) is Anthropic's September 22, 2026 flagship Opus-class reasoning model, succeeding Claude Opus 5 with a 1M-token context window, native image input, adaptive always-on extended thinking, and $4/$20 per million input/output tokens. It leads the frontier on agentic coding, terminal/computer-use, knowledge-work Elo (GDPval-AA v2.1: 1846), and long-horizon tool-use evals, at a reported ~97 tokens/sec on Benchmark Heaven telemetry.

Multi-step deep research investigations

Beam 1.0 is Reflection's open-weight reasoning model for coding and long-horizon agentic workflows.

Ember-1 is Fireworks Research's first custom post-trained model: a Kimi K3 derivative re-trained to emit roughly 40% fewer reasoning tokens at comparable quality, shipping with a 1,048,576-token context window, text+image input, and $3/$15 per-million-token pricing. It is a research preview, not a new architecture — and its only published third-party scores are the Vals AI real-world agentic suites (Vibe Code Bench v1.1 83.9%, CyberBench v1.1 69.5%, Finance Agent v2 51.9%, Terminal-Bench 4.0 ~20.7%).

Glimpse is Elorian's first foundation model — a research-preview "visual thinking" system (released 16 October 2026, with OpenRouter listing on 17 October) that reasons directly over images and video, re-inspecting regions and frames mid-trace rather than converting pixels to a one-shot text description. It ships a 262,144-token context at $0.60/$3.00 per 1M input/output tokens and publishes a compact, vendor-run benchmark card led by CharXiv-R 92.90 and BabyVision 95.36.

Legacy flagship

General-purpose intelligence

Strong general reasoning

PhD-level math, advanced science, security audits, formal proofs

Conversational variant of GPT-5.2

OpenAI's most capable model.

GPT-5.4-mini is a high-performance foundation model engineered by OpenAI, accessible with Zero Data Retention on ARMES.

The speed and throughput tier of the GPT-5.6 family.

The value leader of the GPT-5.6 family.

GPT-6 Sol is OpenAI's mid-tier, cost-efficient high-end GPT-6 model (launched 22 Sept 2026, ~three weeks after flagship GPT-6 Astra), exposing a 1,050,000-token context window, 128K max output, image input, selectable reasoning effort, and hidden test-time reasoning tokens at $2/$10 per 1M tokens — positioned between GPT-6 Astra (quality) and GPT-6 Luna (throughput/price).

GPT-6.1 Sol (openai/gpt-6.1-sol) is OpenAI's September 29, 2026 mid-tier reasoning model, an upgrade to GPT-6 Sol positioned one step below the flagship GPT-6 Astra. It targets agentic coding, computer use, and document-heavy professional automation at roughly one-fifth of Astra's per-token cost, with a 1.05M-token context window, text+image input, and a measured output throughput of ~53 tokens/sec.

SWE-bench Verified: 78.0%, GPQA Diamond: 82.7%, MMLU-Pro: 83.7%, AIME 2025: 90.2%, IFBench: 72.6%, SciCode: 42.0%, Terminal-Bench Hard: 31.0%, τ²-Bench Telecom: 64.4%, HLE: 22.8–30.0%, non-hallucination: ~78% (Thinking mode) | SpaceXSI's large-context multi-agent workhorse (beta Feb 17, 2026; GA Mar 10, 2026).

CaseLaw v2: 79.3% (#1, +25 pts over 4.20), CorpFin: #1, τ²-Bench Telecom: 97.7%, GDPval-AA: 1500 ELO, GPQA Diamond: 90.1%, IFBench: 81.0%, SciCode: 47.3%, Coding Index: 41–42.2%, Terminal-Bench Hard: 38.0%, HLE: 35.0%, Artificial Analysis Intelligence Index: 53 | SpaceXSI's most capable reasoning model (beta Apr 17, 2026; GA Apr 30, 2026).

Artificial Analysis Intelligence Index: 54 (#4 of 168), AA Coding Agent Index: 76 (on par with GPT-5.5), τ³-Banking: 33% (#1 of 28 models), GPQA Diamond: 93%, Terminal-Bench 2.1: 82–83.3%, SWE-Bench Pro: 64.7%, DeepSWE 1.0: 62%, SWE Marathon: 29% (beats Opus 4.8's 26%), ~15,954 output tokens per SWE-Bench Pro task (4.2× more token-efficient than Opus 4.8) | SpaceXSI's flagship model for software engineering, agentic execution, and technical reasoning (released Jul 8, 2026).

AA Intelligence Index: 61 (matches GPT-5.6 Sol), GDPval-AA v2: 1753 Elo, CursorBench v3.2: 69.9%, DeepSWE v1.1: 65.9%, FrontierCode v1.1 Extended: 61.3%, APEX-Agents: 57.5%, Terminal-Bench v3.0: 26%, APEX-SWE: 56.4%, AA-Briefcase: 1577 Elo, Harvey LAB: 15.8%, SWE-bench Verified (Vals): 95.60%, GPQA Diamond: 94.9%, HLE: 42.9%, AA Coding Index: 76.8, AA Agentic Index: 58.7, AA Non-Hallucination Rate: 65.7% | SpaceXSI's current frontier model (released Aug 12, 2026; knowledge cutoff Feb 1, 2026).

Grok 4.7 (x-ai/grok-4.7) is SpaceXAI/SpaceXSI's September 21, 2026 flagship reasoning model, built as a new base model with a long reinforcement-learning run targeting multi-hour coding and agentic knowledge work. It offers a 500K-token context window, text + image input, and configurable reasoning effort (low/medium/high/xhigh), priced at $2/$6 per 1M input/output tokens with a 2x surcharge above 200K-token prompts.

Anthropic's high-speed, frontier-class efficient model delivering 73.3% SWE-bench Verified coding performance at $1/$5 per million tokens with extended thinking capabilities.

Dedicated fast 1K/2K image-to-image editing specialist supporting 1 to 8 reference images with structural and identity preservation.

GDPval-AA v2: 1668 Elo (#3; Fable 5: 1760, Sol: 1748, Opus 4.8: 1600), AA-Briefcase: 1548 Elo (#2; Fable 5: 1583, Sol: 1495), BrowseComp: 91.2% (#1; Sol: 90.4), Terminal-Bench 2.1: 88.3% (#2; Sol: 88.8), DeepSWE: 67.5% (#3; Sol: 73.0, Fable 5: 70.0), FrontierSWE: 81.2% (#2; Fable 5: 86.6), SWE Marathon: 42.0% (#1; Opus 4.8: 40.0), Program Bench: 77.8% (#1; Sol: 77.6), Automation Bench: 30.8% (#1; Sol: 29.7), SpreadsheetBench 2: 34.8% (#1; Fable 5: 34.7), JobBench: 52.9% (#2; Fable 5: 57.4), CharXiv (RQ) w/ tool: 91.3%, Zerobench w/ tool: 41.0%, Kimi Code Bench 2.0 (internal): 72.9%.

Analog 35mm film grain, vintage lens halation, motion blur, and raw anti-AI editorial photography specialist (1K, 0–1 ref).

Fast 1K stylized illustration, anime, manga, watercolor, oil painting, and vector graphic specialist (~12.4s p50, 0–1 ref).

Ultra-budget (--.015) rapid 1K concept explorer for quick silhouettes, moodboards, and rough graphic ideation (0–1 ref).

MAI-Image-2.6 is Microsoft AI's flagship diffusion-based text-to-image and image-to-image editing model (released Aug 2026, Foundry preview Sep 4, 2026), not a text LLM

Microsoft MAI-Image-2.6-Flash is a diffusion / flow-matching image generation and editing model — the speed- and cost-optimized sibling of MAI-Image-2.6. It is NOT an autoregressive LLM or MoE, so LLM benchmarks (SWE-bench, GPQA, MATH-500, LiveCodeBench) do not apply. It generates images ~2.8× faster than GPT-Image-2-Medium at under half the price of the flagship, landing #3 on the Artificial Analysis Image Editing Leaderboard.

Mistral Medium 3.5 is a dense 128B instruction-following model from Mistral AI.

Microsoft: MAI-Voice-2-Flash is a neural speech synthesis model by Microsoft, providing expressive voice rendering for real-time conversation.

Pareto 26.10 (unbiased/pareto-26.10) is Unbiased AI's multimodal composite model — a multi-model ensemble that runs several constituent models on one request and synthesizes a single answer. It offers a 1,048,576-token context window, 131,072-token max output, image input, tool calling, and aggressive $0.80/$3.20 per-million-token pricing, positioned as near-frontier capability at a fraction of flagship cost.

Qwen3.8-2.4T-A95B is Alibaba's open-weight, sparse Mixture-of-Experts frontier model (2.4T total / ~95B active parameters) that serves as the downloadable backbone of the hosted Qwen3.8-Max flagship. It pairs a 1,048,576-token context window with hybrid (Gated DeltaNet + Gated Attention) attention, always-on test-time reasoning, and a strong agentic/coding profile — making it the first open-weight Max-class Qwen, positioned against GPT-6 Astra, Claude Opus 5, and Gemini 3.8 Pro.

Qwen3.8-27B is Alibaba Qwen's open-weight, densely-parameterized vision-language model (≈27.8B params, Apache 2.0, released August 14, 2026). It pairs a 262,144-token native context (YaRN-extensible to 1M) with user-switchable 'flexible thinking' and multimodal image input, delivering frontier-adjacent agentic coding and long-horizon tool-use at single-GPU-friendly scale.

ByteDance's dedicated coding model (released Feb 14, 2026 on Volcano Engine; Jul 30, 2026 on OpenRouter).

Seed 2.1 Turbo is ByteDance Seed's lower-latency, cost-optimized member of the Seed2.1 productivity family — a closed, multimodal (text + image, 262,144-token context) model tuned for coding, tool use, and long-horizon agent workflows at $0.50/$2.50 per 1M tokens.

Fast 4K stylized art, illustration, and multi-reference creative blending specialist (~16.3s p50, 2K/4K, 0–14 refs).

Universal 1K/2K image generation & editing workhorse with crisp short typography, 0–14 free reference images, and transparent PNG preservation.

Primary 4K hero asset generator and multi-panel storyboard engine supporting 2K & 4K output with up to 14 reference images.

Flagship reasoning-driven typography, poster, infographic, hex-palette, and commercial product-staging engine (1K/2K, 0–14 refs).

Sesame: CSM 1B is a neural speech synthesis model by Sesame, providing expressive voice rendering for real-time conversation.

Research queries requiring cited sources

Web-augmented research with live search

Strong all-around coding and analysis

Coding, debugging, knowledge work, financial/legal reasoning, business writing, translation, data analysis, research.

Anthropic's most agentic Sonnet yet (released Jun 30, 2026; knowledge cutoff Jan 2026).

Claude Sonnet 5.5 (anthropic/claude-sonnet-5.5) is Anthropic's Sonnet-class workhorse released September 28, 2026 — a 1M-token-context, image-capable, always-thinking autoregressive model priced at $2/$10 per 1M input/output tokens. It succeeds Sonnet 5 as a direct upgrade, delivering a step-change in agentic coding (Terminal-Bench 4.0 70.6% vs Sonnet 5's 10.3%) and computer use (OSWorld 2.1 80.1%) while running at ~140 tps.

Standard coding and analysis

Agentic coding, complex planning, knowledge-intensive tasks

GLM-5.1 is a high-performance foundation model engineered by Z.ai, accessible with Zero Data Retention on ARMES.

GLM-5.2 (z-ai/glm-5.2) is Z.ai's open-weights flagship reasoning model — a ~743B-parameter sparse Mixture-of-Experts (MoE) with ~39–40B active parameters per token, a 1M-token context window, sparse attention via the novel IndexShare indexer, and up to 5-token MTP speculative decoding. Released June 16, 2026 under MIT license, it is purpose-built for long-horizon agentic workflows and project-level software engineering, landing between Claude Opus-tier and GPT-5.x-tier models on the hardest coding/agentic suites while debuting as the top open-weight model on the Artificial Analysis Intelligence Index.

GLM-5.3 (z-ai/glm-5.3) is Z.ai's flagship text-only reasoning model, built on a 753B-parameter sparse Mixture-of-Experts transformer with DeepSeek Sparse Attention. It is a post-training-only upgrade of the GLM-5.2 base, purpose-engineered for long-horizon software engineering, terminal/CLI agent tasks, and emergent cyber-reasoning, delivering a 1M-token context window at ~70 tokens/sec.

Hy4 preview (tencent/hy4-preview) is Tencent Hy Team's open-weight, 770B-total / 49B-active Mixture-of-Experts flagship (Apache 2.0, released Aug 28, 2026) built for long-horizon coding agents, complex tool-use workflows, and productivity tasks, backed by a 1,048,576-token context window and DeepSeek-style sparse attention.

Inkling is Thinking Machines Lab's first open-weight foundation model — a 975B-parameter sparse Mixture-of-Experts multimodal transformer (41B active parameters, released July 15, 2026) that reasons over text, images, and audio with a controllable "thinking effort" dial. Built for general-purpose reasoning, coding, and agentic/tool-use workloads, it delivers broad, balanced capability with strong open-weights standing (e.g., #1 US open-weights model on Artificial Analysis) at ~210 tokens/sec hosted throughput.

— | Previous generation research and reasoning

GPQA Diamond: 87.6%, SWE-bench: 76.8%, HLE-Full w/ Tools: 50.2% (#1), BrowseComp Swarm: 78.4%, TerminalBench: 50.8% | Research with tools, complex multi-step tasks, agentic workflows, strategic reasoning, document analysis, philosophy, deep conceptual work.

SWE-Bench Pro: 58.6 (leading), GPQA Diamond: 90.5%, HLE-Full w/ Tools: 54.0, MathVision (w/ Python): 93.2, GDPval-AA: 1520 Elo (vs 1309 for K2.5), Hallucination rate: 39% (vs K2.5's 65%) | Long-horizon agentic coding, autonomous engineering, full-stack workflows, persistent 24/7 background agents, multi-step DevOps.

(vendor-reported, in-house suites) Kimi Code Bench v2: 62.0 (+21.8% vs K2.6's 50.9; GPT-5.5 69.0, Opus 4.8 67.4), Program Bench: 53.6 (vs K2.6 48.3), MLS Bench Lite: 35.1 (+31.5% vs K2.6 26.7), Kimi Claw 24/7 Bench: 46.9, MCP Atlas: 76.0 (vs K2.6 69.4), MCP Mark Verified: 81.1 (vs K2.6 72.8) | Cost-efficient long-horizon agentic coding — multi-file refactors, hours-long autonomous engineering, CI/CD, large-codebase analysis across Rust/Go/Python/frontend/DevOps.

SWE-Bench Pro: 57.2%, Artificial Analysis Intelligence Index: 54, Claw-Eval Multimodal: 23.8 (matches Sonnet 4.6), Video-MME: 87.7 (rivals Gemini 3 Pro), HLE: 48.0%, autonomous workflows with 1,000+ tool calls | Complex software engineering, long-horizon agentic tasks, real-time video/audio analysis, multimodal chart interpretation, multi-step automated workflows.

MiMo-V2.6-Pro (xiaomi/mimo-v2.6-pro) is Xiaomi's flagship open-weights, natively omnimodal reasoning model — a ~1.02T-parameter sparse Mixture-of-Experts with ~42B active parameters, a 1M-token context window, and GQA + sliding-window attention. It scored 46.32 on the AA Intelligence Index (top open-source model at release) and is optimized for long-horizon agentic, coding, and cybersecurity workloads at aggressive cost ($0.435/$0.87 per 1M in/out), though it runs deliberately at ~49 tokens/sec.

Mistral Large 4 ("ML4", internally nicknamed "le Chonk") is Mistral AI's 1.05T-parameter natively multimodal sparse MoE flagship, released as a public preview on October 6, 2026, with 52B active parameters, a 1.6B vision encoder, a 1M-token context window, up to 256K output tokens, and a rare strength in cybersecurity/agentic-automation evaluations.

Muse Spark Open (meta/muse-spark-open) is Meta's 816B-total / 42B-active open-weight sparse Mixture-of-Experts reasoning model, purpose-built for long-horizon agentic coding. With a 1,048,576-token context window, native multimodal input (text/image/video/audio), and GQA attention across 62 layers, it trades raw code-writing dominance for best-in-class tool orchestration and codebase comprehension at a ~110 tokens/sec serving rate.

SWE-Bench Verified: 70.7–71.9%, PinchBench: 90.0%, RULER @1M: 94.7%, LiveCodeBench v6: 89.0%, IOI 2025: 570, GPQA (no tools): 87.0%, IFBench: 81.7%, IMOAnswerBench: 88.6–92.3%, Terminal-Bench 2.1: 56.4%, Artificial Analysis Intelligence Index: 48 (highest US open model) | NVIDIA's most capable model (released Jun 4, 2026).

— | Deep reasoning, chain-of-thought tasks

V4 Pro is a high-performance foundation model engineered by DeepSeek, accessible with Zero Data Retention on ARMES.

Terminal-Bench 2.1: 87.9%, DeepSWE: 62.7%, CyberGym: 83.3%, AutomationBench: 31.8%, Toolathlon-Verified: 74.1%, HLE w/ tools: 60.0%, GPQA Diamond: 90.1%, Artificial Analysis Intelligence Index: 53.2, AA Coding Index: 68.8, AA Agentic Index: 49.6 | The official GA release of V4 Pro, superseding the April preview (released Aug 13, 2026).

GLM-5.3-Flash is a native multimodal model from Z.ai.

Legacy cost-efficient

Legacy ultra-lightweight

Cost-efficient general tasks

Ultra-lightweight, fastest OpenAI model

GPT-5.4-nano is a high-performance foundation model engineered by OpenAI, accessible with Zero Data Retention on ARMES.

GPT-6 Luna is OpenAI's fast, cost-efficient multimodal variant in the GPT-6 series, positioned below GPT-6 Sol. It offers a 1,050,000-token context window, image input, a May 18, 2026 knowledge cutoff, and selectable reasoning effort (none→max) at an aggressively low $0.10/$0.50 per 1M tokens — making it a high-volume, latency-sensitive workhorse rather than a frontier reasoning flagship.

Claude Haiku 5.5 is Anthropic's fastest, most capable small model (released Oct 7, 2026), a dense autoregressive transformer with adaptive adjustable-effort thinking, a 1M-token context window, and vision input. Priced at $0.10/$0.50 per 1M input/output tokens, it delivers ~237 tps and strong agentic/coding gains over Haiku 4.5.

GPQA Diamond: 90.4%, HLE: 53.2%, BrowseComp: 84.2% (beats DeepSeek V4-Pro 83.4 & Opus 4.7 79.3; ~ties GPT-5.5 84.4), DeepSearchQA: 91.0, MCP-Atlas: 79.1 (beats Opus 4.7's 77.3), Terminal-Bench 2.1: 71.7% (beats Opus 4.7 & DeepSeek V4-Pro), SWE-bench Verified: 78.0%, GSM8K: 95.37%, MATH: 76.28%, Factual hallucination rate: 5.4% (down from 12.5% in preview), MRCR long-dialogue: 75.1% (up from 42.9%) | Grounded factual Q&A and RAG (trained to answer when grounded and flag missing evidence rather than fabricate), web-browsing/research agents, multi-step tool orchestration, customer-support and long multi-turn dialogue (coreference resolution, constraint tracking across turns), professional document generation, focused coding tasks.

Inkling Small is a high-performance foundation model engineered by Thinking Machines, accessible with Zero Data Retention on ARMES.

LFM2.5-24B-A2B is Liquid AI's open-weight, text-only sparse mixture-of-experts reasoning model (24.0B total / 2.45B active parameters) built for fast, private, on-device and edge-hosted agents that call tools. It is a 40-layer hybrid convolution + grouped-query-attention stack with 64 experts and top-4 routing, a native 131,072-token context, and an always-on reasoning pass before every answer — priced at $0.05/$0.70 per 1M input/output tokens on OpenRouter. Its published evaluations target tool-use and instruction-following rather than frontier knowledge or math, where it is deliberately mid-tier.

Strong creative quality via Behemoth co-distillation | Creative writing, storytelling, business communications, emotional support, brainstorming, marketing copy, brand voice, heartfelt content.

MMLU: 79.6, MMLU-Pro: 74.3, GPQA Diamond: 57.2, MMMU: 73.4, MathVista: 70.7, ChartQA: 88.8, DocVQA: 94.4, MGSM: 90.6 (multilingual), LiveCodeBench: 32.8 | Extreme long-context workloads (multi-document summarization, reasoning over vast codebases, long user-history personalization), native image+text understanding (charts, documents, captioning), multilingual chat across 12 languages, and single-GPU local/commercial deployment.

— | Speed-critical inference, real-time applications, rapid response scenarios

Inception's flagship 260K-context diffusion LLM (dLLM), generating and refining text in parallel at over 1,100 tps for real-time agents, code synthesis, and structured workflows accessible with Zero Data Retention on ARMES.

Mercury 3 (inception/mercury-3) is Inception Labs' flagship diffusion language model (dLLM), released October 19, 2026. It decodes tokens in parallel via iterative denoising rather than strict left-to-right autoregression, pairing a 1M-token context window with a vendor-reported ~1,068 tokens/sec on NVIDIA GPUs and aggressive launch pricing of $0.10/$0.35 per 1M input/output tokens — positioning it as a latency-first alternative to autogregressive speed-tier models like Claude Haiku 5.5 and GPT-6 Luna.

Claw-Eval General: 62.3 — strong everyday performance with full multimodal coverage | Everyday tasks at high throughput, native multimodal Q&A (image/audio/video), cost-efficient see-hear-act workflow

MiMo-V2.6-Flash is Xiaomi's open-weights efficiency tier: a 309B-total / 15B-active sparse Mixture-of-Experts model with a 1M-token context window, hybrid sliding-window + global attention, and multimodal (image) input. Positioned as a low-cost agentic/coding workhorse at $0.14/$0.28 per 1M tokens, it trades raw frontier intelligence for cost efficiency.

— | Previous generation coding and productivity

SWE-bench: 80.2% — rivals Claude Opus and GPT-5.2 | Coding, debugging, refactoring, full-stack development, front-end components, test coverage, technical implementation.

PinchBench: 86.2% (5th overall, within 1.2pts of Opus 4.6), SWE-Pro: 56.22%, GDPval-AA: 1495 Elo (highest open-source), MLE Bench Lite: 66.6%, Terminal Bench 2: 57.0% | Agentic coding, full-project delivery, complex engineering systems, autonomous debugging, production incidents, office productivity.

MiniMax-M3 is a natively multimodal sparse Mixture-of-Experts foundation model (~428B total / ~22–23B active parameters) from MiniMax, built for frontier coding, long-horizon agentic work, and 1M-token context tasks. It combines Grouped-Query Attention with a blockwise MiniMax Sparse Attention scheme, accepts text/image/video input with text output, and runs at a measured ~97 tokens/sec. Additive exact-variant scores verified here include SWE-bench Verified 80.5%, SWE-bench Pro 59.0%, SWE-fficiency 34.8%, KernelBench Hard 28.8%, SAGE 50.6%, and SkillsBench 51.5%.

MiniMax-M3.1-Flash is a low-cost, open-weight sparse MoE multimodal model (23B active / ~428B total params) with a 1,048,576-token context, built for high-throughput agentic, coding, computer-use, and cybersecurity workloads at $0.15/$0.60 per 1M tokens.

Ministral-14b-2512 is a high-performance foundation model engineered by Mistral AI, accessible with Zero Data Retention on ARMES.

GPQA Diamond: 79.23% (82.70% w/ tools), HMMT Feb 2025: 93.67% (94.73% w/ tools), AIME 2025: 90.21%, MMLU-Pro: 83.73%, LiveCodeBench v5: 81.19%, SWE-Bench Verified: 60.47%, SWE-Bench Multilingual: 45.78%, RULER @1M: 91.75% (vs GPT-OSS-120B's 22.30), RULER @256K: 96.30%, AA Intelligence Index: 36 | Agentic reasoning, tool calling, long-context workflows, instruction following, multi-agent orchestration.

SWE-Bench Verified: 51.56% (BF16) / 52.80% (NVFP4), PinchBench: 85.37%, GPQA Diamond: 75.44%, MMLU Pro: 81.94%, Terminal-Bench 2.1: 24.58%, HLE: 11.72%, IFBench (loose): 71.88%, BrowseComp: 36.97%, AA Intelligence Index: 23.6, AA Coding Index: 26.8, AA Agentic Index: 13.8, AA Non-Hallucination Rate: 62.4% | NVIDIA's highest-efficiency model, purpose-built for the execution layer of always-on agents (released Aug 11, 2026).

NVIDIA Nemotron 3.5 Super VL is an open-weight, hybrid Mamba-2/Transformer sparse Mixture-of-Experts vision-language model (120B total / ~12B active parameters) with a C-RADIO/RADIO v2.5-H vision encoder and a native 1,048,576-token (1M) context window. It natively ingests text, images, documents, GUI screenshots, and video, and is tuned for reasoning, coding, tool use, and specialized agentic workflows. Priced aggressively at $0.15/1M input and $0.50/1M output, it targets cost-efficient multimodal autonomy.

General-purpose tasks, balanced reasoning and cost

Ultra-lightweight tasks, highest throughput

GPQA Diamond: 82.4%, SWE-bench: 73.1%, AIME: 93.1% | General chat, simple Q&A, translation, notes retrieval, simple scripts, moderate analysis, standard tasks.

Strong on long-document and routine coding | Cost-effective high-volume tasks, long-document processing, routine coding, local deployment on consumer hardware.

DeepSeek V4 Flash 0731 is a sparse mixture-of-experts model (284B total / 13B active parameters) with a 1M-token context, released as a re-post-trained revision of the V4 Flash preview tuned for coding, reasoning, and agentic workflows. It delivers near-flagship agentic/coding performance at exceptionally low cost.

DeepSeek V4.1 Flash (deepseek/deepseek-v4.1-flash) is DeepSeek's 552B-parameter sparse mixture-of-experts frontier-workhorse (released Sept 10, 2026), the first model built on the company's Causal Encoder-Decoder (CED) architecture with asymmetric activation of ~8B params on input and ~16B on output. Paired with a 1M-token (1,048,576) context window, native image modality, aggressive pricing and ~175 tps hosted throughput, it delivers near-flagship reasoning (GPQA Diamond 90.2, HLE-with-tools 63.9, AA Intelligence Index 39.5) and class-leading terminal/agentic coding (Terminal-Bench 2.1 90.6, DeepSWE v1.1 74.2) at a fraction of Pro-tier cost.

DeepSeek V4.1 Flash 1017 is the October 17, 2026 snapshot of DeepSeek V4.1 Flash — a 552B-parameter sparse Mixture-of-Experts model built on DeepSeek's Causal Encoder-Decoder (CED) architecture, activating only ~8B parameters during prefill and ~16B during decode. It pairs a 1,048,576-token context window with native image input, controllable reasoning effort, MIT-licensed open weights, and an extremely low price ($0.15/1M in, $0.60/1M out), positioning it as a cost-dominant agentic/coding workhorse.