Best AI models for Hermes

Anthropic's most capable generally available model as of April 2026. 70% on CursorBench (vs 58% for Opus 4.6), 98.5% XBOW visual-acuity (vs 54.5%), 3x more resolved production tasks at Rakuten. Catches its own logical faults mid-planning. 3x higher image resolution — 3.75MP vs 1.15MP. Substantially better at multi-session memory and long agentic work. Use it for your most demanding tasks where quality matters more than speed.

#1 on Chatbot Arena (1,505 Elo) — now the top-ranked publicly available model. Also #1 on GPQA Diamond reasoning (94.1%), #1 on ARC-AGI-2 (77.1%), and #1 on creative writing Arena. Exceptional value at $2/$12 per 1M tokens. The 2M context handles entire codebases and book-length documents.

OpenAI's new flagship. Massive long-context leap: recall at 512K–1M tokens jumped from ~21% to 74%, making the million-token window actually usable. ARC-AGI-2 up 11.7 points to 85.0%. Terminal-Bench 2.0 at 82.7% and SWE-bench Pro at 58.6% tie it with Kimi K2.6 at the top of the coding charts. GPQA Diamond at 93.6%. Priced at 2× GPT-5.4 — use GPT-5.4 ($2.50/$15) for cost-sensitive work and GPT-5.5 when you need the full capability jump.

The fast, smart everyday workhorse. 87.5% on GPQA Diamond, 79.6% on SWE-bench Verified, 1,462 Elo on Chatbot Arena. Nearly as capable as Opus at lower cost and higher speed. The right choice for high-volume tasks — daily summaries, coding, research — where you need quality without always paying Opus prices.

The new extreme budget pick. Released April 24, 2026 alongside V4-Pro. 284B total / 13B active MoE — at $0.14 per 1M tokens it costs 20× less than GPT-5.4 while staying in the same conversation. MIT license, 1M token context, self-hostable. DeepSeek V3.2 (its predecessor) retires July 2026. If you're cost-optimizing at scale, nothing at this price point comes close.

Leads EQ-Bench Creative Writing at 1,936 Elo — higher than Opus (1,932) on the benchmark specifically designed to resist sycophancy and measure genuine literary quality. Best voice consistency over long documents: tone, register, and style stay coherent across sessions. At 85% lower cost than Opus, it's the smart pick for high-volume writing — drafts, summaries, long-form content pipelines.

The upgraded heir to Opus 4.6 on literary and instruction-following benchmarks. More direct and opinionated tone than 4.6 — fewer hedges, more conviction. Described as "best model in the world for building dashboards and data-rich interfaces" by Val Town. Raises the bar on professional output quality — interfaces, slides, long-form docs. Best for projects demanding precision and creative depth where spending more per token is worth it.

#1 on Chatbot Arena creative writing Elo (1,487) — human raters prefer it for fiction and blogs — and best for AI-tell avoidance across independent evals. The 2M context and 65K output limit are unmatched for long-form projects: entire chapters, full reports, long narrative arcs. Strong multilingual creative writing in 40+ languages. 12x cheaper on input than Opus.

Solid for structured and commercial writing: technical docs, reports, pitch decks, email campaigns. Excellent instruction-following (~92% IFEval). Ranked ~9th on Arena creative writing — noticeably behind Claude and Gemini for fiction and literary prose, but a strong default when your output needs precise formatting or structured argument flow over stylistic voice.

~1,700 EQ-Bench Creative Writing Elo — roughly 87% of Sonnet's literary quality at 80% lower cost. The budget pick for high-volume content: product descriptions, social copy, blog drafts, and content pipelines where you need coherent writing at scale without paying frontier prices on every call. Upgraded from K2.5 with a larger 262K context window and native image input. API live via Moonshot platform.

Leads GPQA Diamond at 94.1% and ARC-AGI-2 visual reasoning at 77.1% — the highest score on both of the hardest published reasoning benchmarks. Near-perfect AIME 2025 math (98%+). Strong across physics, chemistry, biology, and multi-domain expert knowledge. The 2M context makes it uniquely capable for research requiring both depth and breadth in one pass.

~92% on GPQA Diamond, 41.6% on Humanity's Last Exam, and ~92% on IFEval strict compliance. The best-balanced reasoning model: strong on scientific knowledge, reliable at following precise analytical instructions, and capable across coding and writing tasks simultaneously. The GPT-5.4 Pro variant adds extended reasoning for the genuinely hard problems.

21% fewer errors on OfficeQA Pro document reasoning (Databricks), 13% resolution lift on 93-task coding benchmark (Morph), and 90.9% on BigLaw Bench (Harvey). Now accepts images up to 3.75MP — 3x more than Opus 4.6 — making it the strongest pick for visual research, dense diagrams, and chart analysis. Best for research requiring both depth and 1M-token coherence. Claude Sonnet 4.6 leads GDPval-AA retrieval (1,633 Elo) if throughput matters.

89.8% on GPQA Diamond at just $0.50/$3 per 1M tokens — the best reasoning value by a large margin, outperforming models that cost 20x more. The thinking mode shows its chain-of-thought for auditing. 0.34s time-to-first-token. If you run many hard reasoning calls per day and can't justify frontier pricing, nothing else comes close at this price point.

88.2% on GPQA Diamond with a 1M token context window at a fraction of frontier pricing — a genuine dark-horse for knowledge work. Strong on structured knowledge retrieval and multi-step analytical tasks. The same model that leads Terminal-Bench for coding also performs well on reasoning benchmarks, making it unusually versatile. Available via Alibaba Cloud and OpenRouter.

The best single-GPU open model in 2026. 84.3% on GPQA Diamond, 80.0% on LiveCodeBench v6, 89.2% on AIME 2026 math. Dense architecture (all 31B active every call) gives it consistent quality without the coordination overhead of MoE. Genuinely multimodal — text and images. Runs on an RTX 3090/4090 or an M2/M3 Pro MacBook. Apache 2.0 means you can fine-tune and deploy commercially. The cloud-model quality gap is now thin at this tier.

Released April 22, 2026. 77.2% SWE-bench Verified and 87.8% GPQA Diamond — surpassing the previous Qwen3.6 35B MoE on every major benchmark with a single dense 27B checkpoint. Terminal-Bench 2.0: 59.3%. Multimodal (vision + language), unified thinking/non-thinking mode. Apache 2.0. Runs on a 16 GB Mac Mini or RTX 4090. For now use llama.cpp or Unsloth Studio — Ollama GGUFs don't yet pair the vision projector. Supersedes Qwen3.6 35B released just a week prior.

The strongest battle-tested reasoning model on a single RTX 4090. This is the 32B knowledge-distilled version of the 671B DeepSeek R1 — same chain-of-thought training, fraction of the compute. 62.1% GPQA Diamond, 72.6% AIME 2024, 94.3% MATH-500. Released January 2025 and still the most downloaded reasoning model on Ollama (82M+ pulls) — genuinely proven at scale. MIT license, free to fine-tune. Run it with ollama run deepseek-r1:32b.

One trick nothing else matches: a 10 million token context window — fit entire codebases, books, or months of logs in a single prompt. MoE architecture (109B total, only 17B active per token) keeps inference fast despite the scale. Natively multimodal: text and images. MMLU Pro 74.3%, GPQA Diamond 57.2%, DocVQA 94.4%. Note: Ollama serves it at 128K context by default — the full 10M requires a multi-GPU server build. The 67 GB Q4 download fits a single RTX 4090 (24 GB) but it is tight. Not true open source — the Llama 4 Community License restricts deployment at 700M+ MAU.

The best reasoning model for machines with limited VRAM — an 8 GB GPU or a MacBook with 16 GB RAM. The "plus" variant (Phi-4-reasoning-plus) adds RL training on top of the base SFT and is the one to use: 81.3% AIME 2024 and 68.9% GPQA Diamond, which outperforms the DeepSeek R1 70B distill (5× its size) on both benchmarks. Run it with ollama run phi4-reasoning:14b-plus. Short context (32K, tested to 64K) is the main limit — not suitable for large documents. For math, structured analysis, and code review on laptop hardware, nothing at this weight class comes close.