New AI Models

FLUX.1 Fill [pro]

FLUX.1 Fill [pro] is Black Forest Labs' state-of-the-art inpainting and outpainting model in the FLUX.1 Tools family. It takes an input image, a text description, and a binary mask, then regenerates the masked region or extends the image beyond its original borders with edits that integrate naturally into the surrounding pixels. Black Forest Labs reports that in its public inpainting benchmark the model outperforms all competing methods, including Ideogram 2.0 and the open-source FLUX-Controlnet-Inpainting, making it the leading inpainting model to date. It is served through the BFL API and fits workflows such as product photo retouching, object removal, and canvas expansion.

Gemini 2.0 Flash

Gemini 2.0 Flash is Google's workhorse multimodal model. First shown as an experimental release in December 2024, it reached general availability via the Gemini API in Google AI Studio and Vertex AI on February 5, 2025, and also rolled out to Gemini app users. It accepts text, image, video, and audio input with a 1 million token context window and text output, with image generation and text-to-speech to follow. It can natively call Google Search, code execution, and user-defined functions for tool use. Google positions it as a low-latency, highly efficient model for high-volume, high-frequency tasks at scale, noting that it outperforms 1.5 Pro on key benchmarks at twice the speed, which suits production chat, summarization, and agent workloads.

Gemini 2.0 Flash-Lite

Gemini 2.0 Flash-Lite is Google's most cost-efficient model in the Gemini 2.0 family, released in public preview on February 5, 2025 through the Gemini API in Google AI Studio and Vertex AI. It delivers better quality than 1.5 Flash at the same speed and price, outperforming 1.5 Flash on the majority of benchmarks, and matches 2.0 Flash with a 1 million token context window and multimodal input. Built for high-frequency, straightforward workloads such as classification, summarization, and extraction, Google cites a use case generating relevant one-line captions for around 40,000 unique photos for less than a dollar in the paid tier of AI Studio.

Gemini 2.0 Pro Experimental

Gemini 2.0 Pro Experimental is an experimental model from Google, released on February 5, 2025, and described by Google as its best model yet for coding performance and complex prompts. It offers a 2 million token context window, native tool use with Google Search and code execution, and multimodal input with text output at release. Google states it delivers stronger coding performance, world knowledge, and reasoning than any previous Gemini release. The model was available in Google AI Studio, Vertex AI, and the Gemini app for Gemini Advanced users, and fits workflows such as code generation, long document analysis, and search grounded reasoning.

Phi 4 Mini

Phi 4 Mini is Microsoft's compact 3.8B parameter open model in the Phi-4 family, released in February 2025 under the MIT license. The dense decoder-only Transformer supports a 128K token context window, a 200K vocabulary, and function calling for tool use. Official benchmarks report MMLU 67.3, MMLU-Pro 52.8, GSM8K 88.6, and MATH 64.0, matching or beating some models twice its size on reasoning. It fits math, coding, and reasoning workloads in memory and compute constrained environments, including on-device and edge deployment on Windows, iPhone, Android, and IoT devices.

Phi-4-multimodal-instruct

Phi-4-multimodal-instruct is Microsoft's 5.6B parameter open multimodal model released in February 2025 under the MIT license. A single neural network processes text, image, and audio inputs with a 128K token context, covering speech recognition, speech translation, speech summarization, OCR, chart and document understanding, and function calling. Microsoft reports MMMU 55.1, DocVQA 93.2, and ChartQA 81.4 on vision tasks, plus a 6.14% word error rate that ranked first on the OpenASR leaderboard, surpassing WhisperV3, making it a fit for edge devices and latency bound scenarios.

o3-mini

o3-mini is OpenAI's cost-efficient small reasoning model released on January 31, 2025, delivering strong STEM reasoning in science, math, and coding at the cost and latency targets of o1-mini. Developers can pick among three reasoning effort levels (low, medium, high), and the model supports function calling, Structured Outputs, streaming, and the Batch API, with a 200K token context window, 100K max output tokens, and pricing of $1.10 per million input tokens and $4.40 per million output tokens. Per official evaluations, at medium effort it matches o1 on AIME 2024 competition math, GPQA Diamond science questions, and Codeforces, while at high effort it becomes OpenAI's best performing released model on SWE-bench Verified (61% with an internal tools scaffold) and solves over 32% of FrontierMath research-level problems on the first attempt with a Python tool; A/B tests show replies 24% faster than o1-mini. It suits coding assistants, math and science QA, agents, and other latency-sensitive products that need dependable reasoning at low cost.

Mistral Small 3 24B Base

Mistral Small 3 24B Base is the pretrained base checkpoint of Mistral Small 3, a 24B-parameter open-weight LLM released by Mistral AI on 2025-01-30 under the Apache 2.0 license. It has a 32k context window, a Tekken tokenizer with a 131k vocabulary, and supports dozens of languages. On the official model card the base model scores 80.73 on 5-shot MMLU, 54.37 on MMLU Pro, 80.73 on GSM8K, 69.64 pass@1 on MBPP and 45.98 on MATH, and Mistral AI reports that it rivals models three times larger, such as Llama 3.3 70B. The release ships without instruction tuning, so it is meant as a starting point for the community to run fine-tuning or post-training, for example to build domain-specific experts or reasoning-enhanced variants; the instruction-tuned sibling was released at the same time.

Mistral Small 3 24B Instruct

Mistral Small 3 is a latency-optimized 24B parameter language model released by Mistral AI on January 30, 2025 under the Apache 2.0 license, with both pretrained and instruction-tuned checkpoints. According to the official announcement, it is on par with Llama 3.3 70B Instruct while running more than 3x faster on the same hardware, scoring over 81% on MMLU at around 150 tokens/s. With far fewer layers than competing models of its size class, it targets fast-response conversational assistants, low-latency function calling in agentic workflows, domain-specific fine-tuning, and local inference, running quantized on a single RTX 4090 or a 32GB MacBook as an open alternative to proprietary models like GPT4o-mini.

Qwen-Max

Qwen-Max is the flagship commercial text model on Alibaba Cloud Model Studio (Bailian), positioned as the strongest general-purpose model in the Qwen series and available via Qwen Chat and API. Its January 2025 snapshot, qwen-max-2025-01-25, matches the officially released Qwen2.5-Max, a large-scale MoE model pretrained on more than 20 trillion tokens and post-trained with SFT and RLHF. In official evaluations it outperforms DeepSeek-V3 on Arena-Hard (89.4), LiveBench (62.2), LiveCodeBench (38.7) and GPQA-Diamond (60.1), scores 76.1 on MMLU-Pro, and stays competitive with GPT-4o and Claude-3.5-Sonnet. It offers a 32,768-token context window (30,720 input, 8,192 output) plus function calling and structured output, fitting complex reasoning, coding, math, multilingual and agent workloads where output quality matters most.

Qwen-Plus-2025-01-25

Qwen-Plus-2025-01-25 is a dated snapshot of qwen-plus, the balanced commercial text model in Alibaba Cloud Model Studio's Qwen series, from the Qwen2.5 generation. The official model list describes it as the balanced model of the Qwen family, with quality and speed falling between qwen-Max and qwen-Turbo, making it a fit for moderately complex tasks. Compared with the previous 2025-0112 snapshot, this version delivers an overall upgrade in Chinese and English capabilities, with notable gains in coding, logical reasoning, and multilingual performance. Its reply style is substantially tuned toward human preference, producing more detailed and better-formatted answers on objective queries such as math, logical reasoning, and knowledge QA, while creative writing, JSON format adherence, and role-play receive targeted improvements. The snapshot supports function calling and, per the official page, is expected to be maintained until one month before the next snapshot is released.

DeepSeek R1 Distill Llama 70B

DeepSeek-R1-Distill-Llama-70B is the largest model in the DeepSeek R1 distill series, created by fine-tuning Llama-3.3-70B-Instruct on reasoning samples generated by DeepSeek-R1. The model answers with chain-of-thought output and scores well on math and coding benchmarks in the official evaluation table: 70.0 pass@1 on AIME 2024, 94.5 on MATH-500, 65.2 on GPQA Diamond, 57.5 on LiveCodeBench, and a Codeforces rating of 1633. It suits math problem solving, competitive programming, and other reasoning-heavy workloads where a 70B-class model can be deployed. DeepSeek recommends a temperature of 0.5 to 0.7 (0.6 suggested) and keeping all instructions in the user prompt.

DeepSeek R1 Distill Llama 8B

DeepSeek R1 Distill Llama 8B is a dense reasoning model that DeepSeek open sourced alongside DeepSeek-R1 in January 2025. It is built on the Llama-3.1-8B base model and fine-tuned with supervised fine-tuning on reasoning samples generated by DeepSeek-R1, so the smaller model learns to produce long chain-of-thought outputs on math, code, and general reasoning tasks. In the official distilled model benchmark it reaches 50.4 pass@1 on AIME 2024, 89.1 on MATH-500, 49.0 on GPQA Diamond, 39.6 on LiveCodeBench, and a Codeforces rating of 1205, beating much larger general models such as GPT-4o-0513 and Claude-3.5-Sonnet-1022 on the math tests. DeepSeek recommends a temperature around 0.6, no system prompt, and serving through frameworks like vLLM or SGLang. It suits developers who want strong reasoning quality from a compact model that runs on a single consumer or small cloud GPU.

DeepSeek R1 Distill Qwen 1.5B

DeepSeek-R1-Distill-Qwen-1.5B is the smallest model in the DeepSeek-R1 distill family, fine-tuned from Qwen2.5-Math-1.5B using reasoning samples generated by DeepSeek-R1 (the Qwen distills were trained on 800k curated samples). In the official evaluation it scores 28.9 pass@1 on AIME 2024 and 83.9 on MATH-500, both above GPT-4o-0513 in the same table, alongside 33.8 on GPQA Diamond, 16.9 on LiveCodeBench and a 954 CodeForces rating. Released under the MIT license, it produces chain-of-thought outputs up to 32,768 tokens and, at 1.5B parameters, runs on laptops, phones and other edge devices, which makes it a practical pick for on-device math reasoning, lightweight reasoning apps and distillation research.

DeepSeek R1 Distill Qwen 14B

DeepSeek-R1-Distill-Qwen-14B is a dense reasoning model that DeepSeek open-sourced alongside DeepSeek-R1 in January 2025, one of six distilled checkpoints based on Qwen and Llama. It uses Qwen2.5-14B as the base model and is fine-tuned on 800k samples generated by DeepSeek-R1, transferring the larger model's chain-of-thought reasoning into a 14B footprint. On the official benchmark it reaches 69.7 pass@1 on AIME 2024 and 93.9 on MATH-500, both above OpenAI o1-mini, along with 59.1 on GPQA Diamond, 53.1 on LiveCodeBench and a 1481 Codeforces rating. The weights ship under the MIT license for commercial use, run like a Qwen model on vLLM or SGLang, and suit math and code reasoning workloads that need solid chain-of-thought quality under moderate compute budgets.

DeepSeek R1 Distill Qwen 32B

DeepSeek-R1-Distill-Qwen-32B is one of six dense models distilled from DeepSeek-R1 and open sourced at launch. It fine-tunes the Qwen2.5-32B base model on 800k reasoning samples generated by DeepSeek-R1, transferring chain-of-thought reasoning into a mid-sized model that can be deployed locally. On the official benchmark table it reaches 72.6 pass@1 on AIME 2024, 94.3 on MATH-500, 62.1 on GPQA Diamond and 57.2 on LiveCodeBench, ahead of OpenAI o1-mini on all four, with a Codeforces rating of 1691. Weights support commercial use and run on vLLM or SGLang; DeepSeek recommends a temperature near 0.6 and no system prompt. It suits math, code and general reasoning workloads where a full 671B reasoning model is impractical.

DeepSeek R1 Distill Qwen 7B

DeepSeek-R1-Distill-Qwen-7B is a 7B dense reasoning model that DeepSeek fine-tuned from Qwen2.5-Math-7B with 800k reasoning samples generated by DeepSeek-R1, open-sourced on January 20, 2025 as one of the six distilled models in the R1 family. Despite its small size, it scores 55.5 pass@1 on AIME 2024 and 92.8 on MATH-500 in the official evaluation, ahead of GPT-4o-0513 and the much larger open-source QwQ-32B-Preview on these math benchmarks, with 49.1 on GPQA Diamond and 37.6 on LiveCodeBench. It runs like a standard Qwen model under vLLM or SGLang with a 32,768-token generation window, ships under the MIT license, and works best at a temperature around 0.6 with instructions kept in the user prompt, making it a practical pick for math, code, and education use cases that need chain-of-thought reasoning on limited compute.

DeepSeek R1 Zero

DeepSeek R1 Zero is the research-oriented variant in DeepSeek's first-generation reasoning model release, built by applying large-scale reinforcement learning (RL) directly to the DeepSeek-V3-Base model with no supervised fine-tuning and no cold-start data. According to the official repository, it is the first open research to validate that LLM reasoning capabilities can be incentivized purely through RL: self-verification, reflection, and long chain-of-thought generation emerged naturally during training. The model has 671B total parameters with 37B activated and a 128K context window, and it shows strong reasoning performance, but its outputs suffer from endless repetition, poor readability, and language mixing. It is therefore best suited for research on RL-driven reasoning rather than production use, and the release-quality DeepSeek R1 addresses these weaknesses with cold-start data and multi-stage training.

DeepSeek-R1

DeepSeek-R1 is DeepSeek's first-generation open-source reasoning model, released on 2025-01-20 under the MIT license. Built on the DeepSeek-V3-Base MoE architecture with 671B total parameters, about 37B activated per token and a 128K context window, it was trained with a pipeline of two RL stages plus two SFT stages, after DeepSeek-R1-Zero showed that chain-of-thought reasoning can emerge from pure reinforcement learning. Official benchmarks place it close to OpenAI o1-1217, scoring 79.8 on AIME 2024, 97.3 on MATH-500, a 96.3 Codeforces percentile, 90.8 on MMLU and 65.9 on LiveCodeBench. Served via the OpenAI-compatible API as deepseek-reasoner, it also ships with six distilled dense models from 1.5B to 70B based on Qwen and Llama, fitting math, competitive coding, coding agents and reasoning distillation research.

Kimi-k1.5

Kimi-k1.5 is a multimodal reasoning model from Moonshot AI, introduced in January 2025 alongside the technical report 'Kimi k1.5: Scaling Reinforcement Learning with LLMs'. Trained with reinforcement learning on joint text and vision data, with the RL context window scaled to 128k, it supports two reasoning modes. The long-CoT mode matches OpenAI o1 across benchmarks, scoring 77.5 on AIME, 96.2 on MATH 500, the 94th percentile on Codeforces and 74.9 on MathVista, while the short-CoT mode derived via long2short methods scores 60.8 on AIME, 94.6 on MATH 500 and 47.3 on LiveCodeBench, ahead of GPT-4o and Claude Sonnet 3.5. Notably, these results are achieved without Monte Carlo tree search, value functions or process reward models, which suits math, coding and visual reasoning workloads that need reliable step-by-step inference.