Fine-tuning

Cosmos 3 Nano Policy DROID is a 16B-parameter robotics policy model fine-tuned from Cosmos 3 Nano on the DROID dataset. Given natural language instructions and visual observations from a robot camera (image or video), it generates robot action trajectories (JSON 1D list) for manipulation and control tasks. Compatible with multiple robot embodiments including Franka Panda (single/dual), UR, Google robot, WidowX 250, UMI, and Agibot. Supports 16-400 frame action sequences in various DoF configurations (9D-57D). Intended as a reference implementation for post-training Cosmos 3 Nano on specific robot platforms. The action output modality is represented in prose because the current model schema only has text, vision, video, audio, and related capability flags.

No tracked provider route

Cosmos 3 Super Image2Video is a 64B-parameter fine-tuned variant of Cosmos 3 Super specialized for temporally coherent image-to-video generation. Takes a single image (jpg/png/webp at 256p-720p) plus an optional text prompt (up to 4096 tokens) and outputs MP4 video with 5-400 frames (default 189) at up to 720p, with optional muxed AAC stereo audio at 48kHz. Available via Hugging Face Diffusers and vLLM-Omni.

No tracked provider route

Cosmos 3 Super Text2Image is a 64B-parameter fine-tuned variant of Cosmos 3 Super specialized for high-fidelity text-to-image generation. Takes text prompts up to 4096 tokens and outputs JPEG images at 256p, 480p, or 720p in aspect ratios 16:9, 4:3, 1:1, 3:4, or 9:16. Available via Hugging Face Diffusers (DiffusionPipeline) and vLLM-Omni.

No tracked provider route

OpenAI GPT-4o: Flagship multimodal model with vision, function calling, and broad capability. $2.50/M input, $10/M output.

$2.50 in / $10.00 out / 1M tokens

5 routes · 1 batch · 3 cache

OpenAI: GPT-4o-mini available via OpenRouter. Pricing: $0.15/1M input, $0.6/1M output.

$0.150 in / $0.600 out / 1M tokens

4 routes · 3 cache

Inkling is Thinking Machines Lab's Apache-2.0 open-weight general-purpose multimodal model. It accepts text, image, and audio inputs and generates text with a 1M-token context window.

$1.00 in / $4.05 out / 1M tokens

3 routes · 3 cache

Tencent Hunyuan Hy4 preview is a Mixture-of-Experts flagship language model released and open-sourced August 28, 2026. First-party Tencent release: 770B total parameters with 49B activated per token, context window exceeding 1M tokens, aimed at coding, office productivity, game development, and scientific research. Hugging Face model card (tencent/Hy4-preview) documents a 78-layer MoE backbone (256 routed experts, top-8 plus 1 shared expert), Gated DeepSeek Sparse Attention with IndexCache, iHC residual streams, native MTP speculative decoding, Apache License 2.0 weights (also Hy4-preview-FP8), and OpenAI-compatible serving via vLLM/SGLang with tool-call and reasoning parsers. Public access: Hugging Face/ModelScope/GitCode/CNB weights; API via Tencent Cloud TokenHub and OpenRouter (model id tencent/hy4-preview); also WorkBuddy, CodeBuddy, Yuanbao, and ima. Official API list pricing USD 0.834 / 2.501 per 1M input/output tokens and USD 0.042 per 1M cache hits.

$0.7506 in / $2.25 out / 1M tokens

2 routes · 2 cache

GLiNER2.5-Decide is Fastino Labs' 340M open-weight encoder decision and classification model in the GLiNER2.5 family, released September 24, 2026. The Hugging Face card (Apache 2.0) describes a DeBERTa-v3-large encoder with base model fastino/gliner2-large-v1. It loads via gliner2 and AutoExtractor.from_pretrained("fastino/GLiNER2.5-Decide"), accepts any label set at call time (intent, routing, sentiment, priority, policy, multi-label) in one forward pass, and returns structured decisions with probabilities and confidence. It is not a chat, reasoning, or open-QA model. Hosted inference and fine-tuning use the Fastino API model ID fastino/gliner2.5-decide. The card's Fast Decisions suite (average 60.2%) is an internal evaluation and is not recorded as a catalog benchmark.

$0.030 in / Free out / 1M tokens

1 route

Laya is ConvAI Innovations' English open-weight System 1 decision model (ModernBERT-large, 421M, context 512), released on Hugging Face on 2026-09-18 under Apache 2.0. It is non-autoregressive: a state plus typed questions (choice, yes/no (called "noul" in the API) or score) returns calibrated probabilities in one forward pass, trained with RLCD, and it does not generate text. Install with pip install laya; the Router sends English inputs here and other languages to Laya Multilingual. The family hub also ships laya-multilingual and laya-typed-decisions. A community package, laya-serve, exposes a Jev-compatible POST /v1/systemone endpoint; ConvAI does not publish a hosted per-token price.

$0.0105 in / Free out / 1M tokens

1 route

Xing4.0-29B-A4B is an open-weight mixture-of-experts language model from China Telecom Artificial Intelligence Technology Co., Ltd. (TeleAI), the successor to the TeleChat series. It has 29B total parameters with 4B active per token (64 routed experts, 4 active, 1 shared), MLA attention, and a native 256K context (extensible to 512K). Weights are published under Apache 2.0 at Hugging Face XingChen-AGI/Xing4.0-29B-A4B. Open-sourced 2026-09-17.

Pricing not tracked / 1M tokens

1 route

d1-3B is Liquid AI's open-weight 3.12B multimodal decision model (post-trained from LFM2.5-VL-3B). It takes a state (text, JSON, images, or a mix) plus named questions of type choice, yes/no (called "noul" in the API) or score and returns answers in one forward pass with zero output tokens. Context length 32,768. Released 2026-10-07 on Hugging Face under the LFM Open License v1.0 (free commercial use below $10M annual revenue). Not a chat model. Liquid reports 8 ms/decision on an RTX 4090. Hosted proprietary d1 API remains a separate catalog entry.

No tracked provider route

d1-omni-600M is Liquid AI's open-weight ~587M edge decision model (post-trained from LFM2.5-Encoder-350M) with text+image or text+audio inputs. It returns typed choice, yes/no (called "noul" in the API) or score answers with zero output tokens. Context 16,384 tokens across modalities (with images, state+question text is cut to 896 tokens as trained). Audio path: up to 30 s speech; trained on English speaker↔assistant requests. Released 2026-10-07 under LFM Open License v1.0. Liquid calls it an experimental checkpoint. Not a chat model.

No tracked provider route

Laya Multilingual is the 322M-parameter, mmBERT-based member of the Laya family: a non-autoregressive decision model that takes a text, email, ticket, or JSON input plus typed questions and returns typed answers with probabilities in one forward pass, across 100+ languages. It ships with a 1,024-token limit but reads up to 8,192 tokens when called with `max_len=8192`, and it is the Laya checkpoint for anything that is not English.

No tracked provider route

Laya Typed-Decisions is the ModernBERT-large (421M, context 1024) fine-tuned Laya checkpoint for typed-decisions workflows such as agent-trace observability, customer service, invoice processing, and security incidents. It was created on Hugging Face on 2026-09-18 under Apache 2.0. The card reports 0.766 accuracy on that suite; that custom evaluation is not a catalog benchmark.

No tracked provider route

Needle 3 is Cactus Compute's sliceable automation foundation model for tiny devices, announced 2026-09-17. One set of weights; every depth from 2 to 20 layers is its own deployable model (intelligence ladder). Params ~25–121M at CQ2-bit (2L ~25M, 4L ~29M, 8L ~52M, 16L ~98M, 20L ~121M / config total_parameters 121021910); footprint 8–29MB (.cact; full 20L ~29MB). Positioning: does not chat — every turn is tool calling or schema extraction (structured JSON); also returns text embeddings. First-party claims it can match DeepSeek V4 Flash (esp. after fine-tune from 4L up on downstream tools) and beat much larger models on mobile tool calls / match 2–3× larger on extraction. Trained on 360B tokens of proprietary structured data. Context 8192 (HF config max_position_embeddings +.cact-format blog). Apache 2.0 (HF license:apache-2.0 + LICENSE file). Weights https://huggingface.co/Cactus-Compute/needle3 (needle3.cact + engines); pip cactus-needle; GitHub cactus-compute/needle.

No tracked provider route

Spark-X2.5-1.7B is the smaller SparkLLM / XHToken (iFlytek) post-trained on-device general language model in the Spark-X2.5 series. Hugging Face card XHToken/Spark-X2.5-1.7B (Apache License 2.0 LICENSE blob, Copyright 2026 XHToken) documents Spark2_5ForCausalLM with the same hybrid 3:1 sliding-window / full-attention pattern, native 1,048,576-token context, safetensors 1,707,657,216 parameters (card title 1.7B; HF size widget rounds to 2B), thinking mode, tool-call parser spark25, and Llama-Factory fine-tuning. Same first-party GitHub README as the 4B: ~20T pretrain, SFT plus RL / MOPD from base XHToken/Spark-X2.5-1.7B-Base; 200+ languages; agent/coding/reasoning on-device. Serving: SGLang, vLLM, MLX, llama.cpp, Ollama, LM Studio. Xingchen MaaS listed on SparkLLM DEV as limited-time free; first-party Token Plan docs on 2026-09-01 list Spark-X2 / X2-Agent / X2-Flash only, not X2.5 IDs or USD rates. Sibling Spark-X2.5-4B.

No tracked provider route

Spark-X2.5-4B is SparkLLM / XHToken (iFlytek) post-trained on-device general language model in the Spark-X2.5 series. Hugging Face card XHToken/Spark-X2.5-4B (Apache License 2.0 LICENSE blob, Copyright 2026 XHToken) documents a compact dense decoder (Spark2_5ForCausalLM) with hybrid attention: repeating 3 sliding-window layers plus 1 full-attention layer, native context up to 1,048,576 tokens (config max_position_embeddings), safetensors 4,112,079,360 parameters, thinking-mode chat template, tool-call parser spark25, and Llama-Factory fine-tuning. First-party GitHub README and HF card: pretrained ~20T tokens then SFT plus large-scale RL / MOPD from base XHToken/Spark-X2.5-4B-Base; 200+ languages; conversation, writing, translation, reasoning, coding, tool use, and agent harnesses (Codex, Claude Code, OpenClaw, Hermes). Serving recipes for SGLang, vLLM, MLX, llama.cpp, Ollama, LM Studio. Weights also listed on ModelScope, Modelers, SCNet. Xingchen MaaS API mentioned on SparkLLM DEV as limited-time free without published USD list prices or X2.5 providerModelId on first-party Token Plan as of 2026-09-01. Sibling Spark-X2.5-1.7B.

No tracked provider route

Needle 2 is Cactus Compute's prior-generation open 45M-parameter automation model (first-party blog 2026-08-09 + HF Cactus-Compute/needle2). Single ~14MB CQ2-bit binary; ~28MB session RAM; tool calling, device use, and structured extraction. Apache 2.0 (HF license:apache-2.0). Tip Needle 3 is the follow-up under the same Needle family. Blog emphasizes 256-token sliding KV window for RAM bounding rather than publishing a clear 8k-style context figure.

No tracked provider route

Cohere Transcribe Arabic is Cohere's named 2B Arabic/English automatic speech-recognition model: an audio-in, text-out fine-tune of Cohere Transcribe for Arabic dialects and Arabic-English code-switching. It is not a general text-generation model.

No tracked provider route

NVIDIA Nemotron-Labs-Diffusion 14B is the largest text model in NVIDIA Research's diffusion language model family, released May 23, 2026. Uses diffusion-based parallel decoding enabling up to 6× higher throughput versus autoregressive baselines, with three decoding modes: autoregressive, diffusion, and self-speculation. Training code released through NVIDIA Megatron Bridge framework alongside the weights, enabling fine-tuning. Released under NVIDIA Nemotron Open Model License (commercially usable open weights). Available on Hugging Face at nvidia/Nemotron-Labs-Diffusion-14B.

No tracked provider route

Original Needle (catalog slug needle-1) is Cactus Compute's first open ~26M-parameter function-calling model (first-party blog 2026-05-12 + HF Cactus-Compute/needle). Distilled Gemini tool-calling into a Simple Attention Network for consumer devices. License note: May 2026 blog states Apache 2.0; HF cardData/tags currently license:mit and no LICENSE sibling file License: HF cardData/tags mit (no LICENSE blob in tree); May 2026 product blog also stated Apache 2.0

No tracked provider route