0.8.8

Added

• Kimi K2.7 Code (Moonshot AI, China) — open-source, coding-focused model in the Kimi K2 family, built for reliable end-to-end programming over long contexts. 1-trillion-parameter model that cuts reasoning token usage ~30% vs K2.6 while improving coding and agent performance (+21.8% Kimi Code Bench v2, +11.0% Program Bench, +31.5% MLS Bench Lite). Modified MIT License, 1M context window.
• 3 new mappings:
  • Kimi K2.7 Code on Moonshot ($1.00/$3.00 per 1M tokens)
  • Kimi K2.7 Code on Hugging Face Inference ($0.60/$2.40 per 1M tokens)
  • Kimi K2.7 Code on OpenRouter ($0.60/$2.40 per 1M tokens)

0.8.7

Added

• Claude Fable 5 (Anthropic, US) — first publicly available Mythos-class model, exceeding any model Anthropic has previously made generally available. State-of-the-art on nearly all tested benchmarks with exceptional software engineering, knowledge work, vision, and scientific research; its lead grows on longer, more complex tasks. Ships with safeguards that route sensitive cybersecurity, biology, chemistry, and distillation queries to Opus 4.8. 300K context window.
• Claude Mythos 5 (Anthropic, US) — the same underlying model as Fable 5 with safeguards lifted in some areas. Strongest cybersecurity capabilities of any model in the world. Access restricted to trusted cyberdefenders and infrastructure providers via Project Glasswing. 300K context window.
• DiffusionGemma (Google, US) — experimental diffusion-based member of the Gemma 4 open family. Denoises a canvas of placeholder tokens to generate up to 256 tokens in parallel rather than autoregressively, delivering ~4x throughput of similarly sized Gemma models on local hardware. MoE with 26B total / 3.8B active parameters, Apache 2.0, 256K context window.

0.8.0

Added

• 16 new provider-model mappings (total: 151)

Changed

• meta/muse-spark mapping: regions updated from us-east-1, us-west-2 to global
• alibaba-model-studio provider: added ap-east-1 (Hong Kong) region

Improved

• validate_registry.py: added referential integrity check — mapping available_regions must exist in the provider's declared regions
• validate_registry.py: global in provider regions acts as wildcard, allowing any region in mappings

0.7.8

Added

• 9 new models:

  • Jamba Large 1.7 (AI21 Labs, Israel) — hybrid SSM-Transformer MoE, 256K context, enterprise-grade
  • Yi-Lightning (01.AI, China) — MoE architecture, top Chatbot Arena in Chinese/Math/Code
  • Falcon-H1 (TII, UAE) — hybrid Mamba-Transformer, outperforms Llama/Qwen in 30-70B range
  • Falcon 3 10B (TII, UAE) — #1 on HuggingFace leaderboard under 13B params
  • Palmyra X5 (Writer, USA) — 1M context window, adaptive reasoning, enterprise agents
  • DBRX (Databricks, USA) — 132B MoE (36B active), open-source enterprise model
  • Snowflake Arctic (Snowflake, USA) — 480B MoE (17B active), Apache 2.0, SQL/code specialist
  • StableLM 2 12B (Stability AI, UK) — 12.1B decoder, 2T tokens multilingual training
  • Alloma 8B Instruct (Uzbek LLM Lab, Uzbekistan) — first Uzbek-optimized LLM with custom tokenizer

• 5 new providers:

  • AI21 Labs — Jamba model family with hybrid SSM-Transformer architecture
  • Reka AI — multimodal models (text/image/video/audio) with Flash and Edge variants
  • Lambda — GPU cloud and managed inference API for open-source models
  • Snowflake Cortex AI — Arctic models integrated with Snowflake data platform
  • 01.AI — Yi model family with strong Chinese/multilingual capabilities

• New countries represented: Israel (IL), UAE (AE), Uzbekistan (UZ), UK (GB)

0.7.7

Added

• Country of origin field (country) added to all 88 models using ISO 3166-1 alpha-2 codes
• Countries represented: US, CN, FR, KZ, KR, AE, RU
• Model schema updated with optional country field

0.7.6

Added

• Solar Pro 3 — Upstage's 102B MoE language model (12B active params) with 128K context. Optimized for Korean with English and Japanese support. Strong reasoning, structured output, and agentic workflows.
• K2 Think — LLM360/MBZUAI's 32B open-weights reasoning model (Apache 2.0). Trained with RL and verifiable rewards for math, science, and code. ~2000 tok/s on Cerebras WSE.
• New provider: Upstage — Korean AI company with OpenAI-compatible API

0.7.5

Added

• Qwen 3.7 Max — Alibaba's flagship proprietary model for advanced agentic coding, complex reasoning, and long-horizon task execution. Ranked #13 in Arena AI Text, #7 in Math, #10 in Coding. Supports 1000+ tool integrations and 35-hour sustained autonomous operation.
• Qwen 3.7 Plus — Alibaba's multimodal variant optimized for vision understanding. Ranked #5 globally in Arena AI Vision leaderboard.

0.7.4

Added

• Gemini 3 Flash — Google's balanced model combining Gemini 3 Pro reasoning with Flash-line latency and cost efficiency. 1M context, configurable thinking levels, streaming function calling.
• Gemini 3.1 Flash-Lite — Google's most cost-efficient model optimized for high-volume, low-latency tasks. 2.5x faster TTFT vs Gemini 2.5 Flash, 1M context, full multimodal support.
• Mappings: Gemini 3 Flash on Google Vertex AI and Google AI Studio ($0.50/$3.00 per 1M tokens)
• Mappings: Gemini 3.1 Flash-Lite on Google Vertex AI and Google AI Studio ($0.25/$1.50 per 1M tokens)

0.7.3

Added

• MiniCPM-V 4.6 — OpenBMB's ultra-efficient 1B multimodal model (vision + video), edge-deployable, 256K context
• Aya Expanse 32B — Cohere For AI's 32B multilingual model, 23 languages, 8K context
• Tiny Aya — Cohere For AI's compact 3.35B multilingual model, 70+ languages, edge-optimized

0.7.2

Added

• Hy3 Preview — Tencent's 295B MoE / 21B active, fast+slow thinking, 256K context, open-weight
• Laguna M.1 — Poolside AI's 225B MoE / 23B active, agentic coding flagship, 128K context
• Mappings: Hy3 Preview on SiliconFlow and OpenRouter, Laguna M.1 on OpenRouter