Available models¶
Blablador hosts a rotating set of models. This page lists the models currently available on the service, together with all of their usable names (model values for the API and the model dropdown in the web interface).
You can always get the live, up-to-date list with:
The list below reflects the output of that endpoint as of August 2026. Each root model is published under one or more names: a descriptive name (what you see in the Blablador web interface) and one or more stable aliases. Use an alias in your code so that the underlying model can change without breaking it.
Overview¶
| Model | Context window | Stable aliases |
|---|---|---|
| Kimi-K3-1M | 1,048,576 tokens | alias-kimi-k3-1m |
| GLM-5.2-AWQ-INT4 | 1,048,576 tokens | alias-glm-huge |
| DeepSeek-V4-Flash-0731 | 1,048,576 tokens | alias-deepseek-v4-flash-0731 |
| Muse Glimmer 30B | — | alias-muse |
| MiniMax-H3 | — | (none) |
| Qwen3.8-27B-FP8 | 262,144 tokens | alias-qwen38-27b, alias-code |
| Qwen3.5-122B-A10B-FP8 | 262,144 tokens | alias-large |
| MiniMax-M2.7 | 131,072 tokens | alias-huge |
| GPT-OSS-120B | 131,072 tokens | alias-fast |
| EVE-Instruct | 131,072 tokens | alias-eve |
| Apertus-8B-Instruct-2509 | 65,536 tokens | alias-apertus |
| Qwen3.8-27B (DFlash) | 217,872 tokens | alias-qwen3.8-27B-dflash |
| Qwen3-Embedding-8B | 8,192 tokens | alias-embeddings, alias-qwen3-8b-embeddings, gpt-3.5-turbo, text-davinci-003, text-embedding-ada-002 |
| faster-whisper-large-v3 | — | (none) |
Frontier models (1M context)¶
These three models share the 1-million-token context ceiling and represent the current top of the capability hierarchy.
Kimi-K3-1M¶
Moonshot AI's current flagship (August 2026): an open-weight 2.8T-parameter native multimodal agentic model, the world's first open 3T-class model. It is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) with a Stable LatentMoE framework that activates 16 out of 896 experts for roughly 2.5× scaling efficiency over K2. It ships with native vision, a 1M-token context window, and is designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning — which makes it the best model on the service at the cost of being the slowest.
99 - Kimi-K3-1M - our best (and slowest) model as of August 2026alias-kimi-k3-1m
GLM-5.2-AWQ-INT4¶
Zhipu AI's open flagship (MIT licence), a sparse MoE model with solid 1M-token context for stable long-horizon work. GLM-5.2 introduces the IndexShare technique (reusing the same indexer across every four sparse attention layers, cutting per-token FLOPs by 2.9× at 1M context) and an improved MTP (Multi-Token Prediction) layer that raises speculative-decoding acceptance length by up to 20%. It also ships with multiple coding effort levels to balance performance and latency.
99 - GLM-5.2-AWQ-INT4alias-glm-huge
DeepSeek-V4-Flash-0731¶
DeepSeek's preview of the V4 series (284B total parameters, 13B activated, 1M context) released in July 2024. The Flash variant uses a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), which at 1M-token context reduces FLOPs to 27% of single-token inference and KV cache to 10% compared with DeepSeek-V3.2 — making the long context practically usable.
DeepSeek-V4-Flash-0731alias-deepseek-v4-flash-0731
Vision and agentic models¶
Muse Glimmer 30B¶
Meta Superintelligence Lab's August 2026 model: a 30B-parameter causal language model with a dedicated perception encoder, distilled from Muse Spark for autonomous agentic tasks on consumer hardware. It integrates multi-step reasoning, reliable tool use, multimodal understanding, and failure recovery into a single model that runs locally without cloud access. Strong on DeepSearch QA, MCP-Atlas, τ3-Bench, and SWE-Bench.
10 - Muse Glimmer 30b - the newest META model as of August 11, 2026alias-muse
Qwen3.8-27B-FP8¶
A compact dense multimodal model from the Qwen3.8 family (August 2026) that understands images and video, with flexible thinking control. It is a good all-round workhorse for coding, professional work, research, and multi-step agentic tasks, in FP8-quantized form with a 256k context window. Note that this model is currently also the target of the alias-code alias.
07 - Qwen3.8-27B-FP8 - Multimodal model from Aug 2026alias-qwen38-27balias-code
General-purpose models¶
Qwen3.5-122B-A10B-FP8¶
A large general-purpose model from the Qwen3.5 family: a 122B-parameter sparse mixture-of-experts with about 10B activated parameters, trained with early-fusion multimodality and strong reinforcement learning. The FP8-quantized checkpoint runs at high throughput with minimal latency overhead, and it is our recommended default for a good accuracy/speed balance (alias-large).
02 - Qwen3.5-122B-A10B-FP8, general purpose large modelalias-large
MiniMax-M2.7¶
MiniMax's flagship model (April 2026), strongly oriented towards agentic work: building complex agent harnesses, multi-step productivity tasks, and real-world software engineering (bug troubleshooting, refactoring, SRE-level reasoning). It scores near the top of the MLE-Bench ML-competition benchmark and initiates a cycle of model self-evolution during its own training.
01 - MiniMax-M2.7 - our best model as of April, 2026alias-huge
GPT-OSS-120B¶
OpenAI's open-weight 120B model (released in August 2025), served here in its MXFP4-quantized form. It is a sparse mixture-of-experts model with about 5.1B activated parameters, which gives it high throughput for its quality level — which is why it currently serves as our fast alias.
01 - GPT-OSS-120b - an open model released by OpenAI in August 2025alias-fast
EVE-Instruct¶
ESA's domain model for Earth Intelligence, a fine-tune of Mistral-Small (24B, 128k context) specialized in Earth Observation and Earth Science. It was adapted with an instruction-fine-tuning plus long-form-text mixture (~33.5B tokens, with a published subset of ~10.7B) that boosts EO/ES capabilities while preserving the base model's general conversational and tool-use abilities.
20 - EVE-Instruct - Expert Earth Observation and Earth Science (ES) domainsalias-eve
Apertus-8B-Instruct-2509¶
A Swiss open model from the Swiss National AI Institute (SNAI, the ETH Zurich / EPFL partnership), released in September 2025. Apertus is a fully-open, multilingual model (1,000+ languages, long context) trained exclusively on compliant, fully open data, and the 8B instruct version here is the smaller sibling of the 70B original.
15 - Apertus-8B-Instruct-2509 - A new swiss model from September 2025alias-apertus
Qwen3.8-27B (DFlash)¶
The same Qwen3.8-27B model as above, served on our haicluster node with DFlash speculative decoding for faster inference. It is a good small model of August 2026 when you want speed; the context window is reduced (~218k tokens) to fit the DFlash draft model.
10 - Qwen 3.8-27B with DFlash on haicluster. A good small model of August 2026alias-qwen3.8-27B-dflash
Media generation¶
MiniMax-H3¶
MiniMax's omni-modal generative system (the model behind Hailuo), which unifies understanding and generation across text, images, video, and audio. It generates video with native stereo audio at resolutions up to 2K and durations of up to 15 seconds, and supports prompt-to-video, image-to-video, video editing, and audio-visual synthesis. As a generation model it does not report a context window, and it is currently served without a stable alias — use its full path as-is.
/p/project1/atmlaml/strube1/FastChat/models/MiniMax-H3
Embeddings¶
Qwen3-Embedding-8B¶
Qwen's text-embedding model (6,000+ dimensions, 100+ languages, 32k input), used for RAG, retrieval, and similarity search — i.e. for the /v1/embeddings endpoint rather than chat. It is reachable under its alias as well as under three OpenAI-style names, so that frameworks with hardcoded OpenAI model names (LangChain and friends) work out of the box.
alias-embeddingsalias-qwen3-8b-embeddingsgpt-3.5-turbotext-davinci-003text-embedding-ada-002
Speech recognition¶
faster-whisper-large-v3¶
A CTranslate2-optimized build of OpenAI's Whisper large-v3 automatic speech recognition model, i.e. speech-to-text rather than chat. It supports a large number of languages (English, German, French, Spanish, Chinese, and many more) and is the only model on the service without an alias, so use its name as-is.
faster-whisper-large-v3
Models come and go, so this list is a snapshot: treat the API /v1/models endpoint as the source of truth. The alias → model mapping (and the current targets of alias-fast, alias-large, etc.) is also maintained in Blablador API access.