{"umans-mimo-v2.6-pro-lab":{"display_name":"Umans MiMo V2.6 Pro (lab)","provider":"Xiaomi","retired_on":"2026-09-24","stage":"playground","note":"playground experiment; the MiMo V2.6 Pro lab window ran Sep 22 to 24, 2026","description":"MiMo V2.6 Pro as a Labs experiment, open for a short test window: temporary, not a permanent id. Xiaomi's flagship open model, served from the official MiMo-V2.6-Pro-RL open weights: a 1.02T mixture-of-experts (42B active per token) built for agentic coding and long-horizon tool use, with native image, video and audio understanding and a 1M-token context. Thinking is on by default; the low, medium and high reasoning dials all enable the same full thinking today (fine-grained effort is not differentiated yet), and none turns thinking off. Access is seat-gated through the Labs page while an experiment is live. It is offered at limited capacity and low availability, so expect it to be flaky and to go down under load: crash it, give it a moment, and try again.","weights":{"precision":"fp8","hf_url":"https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL"}},"umans-deepseek-v4-pro-0813":{"display_name":"Umans DeepSeek V4 Pro","provider":"DeepSeek","retired_on":"2026-09-14","replacement":"umans-deepseek-v4.1-flash","note":"no longer available; use umans-deepseek-v4.1-flash","description":"DeepSeek V4 Pro, served from the official 0813 release: DeepSeek's flagship coding and reasoning MoE, built for long-horizon agentic coding and demanding tool-heavy workloads on a 1M-token context window. The 0813 release supersedes the April preview with substantially stronger agentic performance. Reasoning has three modes: non-think (none), think high (high, the default) and think max (max). Billed per token ($1.32 / $3.96 / $0.044 per 1M; input / output / cache read). Served on our own GPU infrastructure with high availability. Deprecated: use `umans-deepseek-v4.1-flash` instead (sunset 2026-09-14).","weights":{"precision":"full","hf_url":"https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813"}},"umans-glm-5.2":{"display_name":"Umans GLM 5.2","provider":"GLM","retired_on":"2026-09-10","replacement":"umans-glm-5.3","note":"no longer available; use umans-glm-5.3","description":"GLM 5.2 is our best model for coding right now, with a 400K context window for large codebases. Vision is available on the Anthropic Messages API (`/v1/messages`) only, through a server-side handoff (GLM 5.2 generates the text, Kimi preprocesses the image); that handoff will be retired soon in favour of more efficient client-side image handling. Deprecated: use `umans-glm-5.3` instead (sunset 2026-09-10).","weights":{"precision":"fp8","hf_url":"https://huggingface.co/zai-org/GLM-5.2-FP8"}},"umans-kimi-k2.7":{"display_name":"Umans Kimi K2.7 Code","provider":"Moonshot","retired_on":"2026-09-10","replacement":"umans-kimi-k3","note":"no longer available; use umans-kimi-k3","description":"Kimi K2.7-Code via Umans Code - Moonshot's strongest coding model and the successor to Kimi K2.6. Built for complex, tool-heavy agentic coding; it reasons more efficiently than K2.6, so agent sessions run faster at the same depth. Deprecated: use `umans-kimi-k3` instead (sunset 2026-09-10).","weights":{"precision":"full","hf_url":"https://huggingface.co/moonshotai/Kimi-K2.7-Code"}},"umans-glm-5.3-flash-lab":{"display_name":"Umans GLM 5.3 Flash (lab)","provider":"GLM","retired_on":"2026-09-08","replacement":"umans-glm-5.3-flash","stage":"playground","note":"playground experiment; the GLM-5.3-Flash pre-release window ran Aug 26 to Sep 8, 2026; its metrics carry on the released model's status page as its pre-release period","description":"GLM-5.3-Flash as a Labs experiment, open for a short test window: temporary, not a permanent id. Served from the native weights (Z.ai's GLM-5.3-Flash): a 320B mixture-of-experts model with 18B active parameters per token, the first natively multimodal release in the GLM-5 series, built for fast coding and agentic tasks on a 1M context window. It always thinks at max effort by default; dial reasoning to low or high (thinking cannot be turned off). Access is seat-gated through the Labs page while an experiment is live. It is offered at limited capacity and low availability, so expect it to be flaky and to go down under load: crash it, give it a moment, and try again. For production work we recommend umans-coder or umans-deepseek-v4-pro-0813.","weights":{"precision":"fp8","hf_url":"https://huggingface.co/zai-org/GLM-5.3-Flash"}},"umans-deepseek-v4-flash-vision-exp-lab":{"display_name":"Umans DeepSeek V4 Flash Vision Exp (lab)","provider":"DeepSeek","retired_on":"2026-09-05","stage":"playground","note":"playground experiment; the DeepSeek V4 Flash Vision Exp lab window ran Sep 1 to 5, 2026","description":"DeepSeek V4 Flash Vision Exp as a Labs experiment, open for a short test window: temporary, not a permanent id. DeepSeek's first multimodal V4 model, served from the official DeepSeek-V4-Flash-Vision-Exp open weights: the fast agentic coding MoE (284B total, 13B active) with native image understanding, on a 1M-token context. Reasoning has four modes: non-think (none), think low (low, the default), think high (high) and think max (max). Access is seat-gated through the Labs page while an experiment is live. It is offered at limited capacity and low availability, so expect it to be flaky and to go down under load: crash it, give it a moment, and try again. For production work we recommend umans-coder or umans-deepseek-v4-pro-0813.","weights":{"precision":"fp8","hf_url":"https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp"}},"umans-qwen3.8-27b-lab":{"display_name":"Umans Qwen3.8 27B (lab)","provider":"Qwen","retired_on":"2026-08-21","stage":"playground","note":"playground experiment; the Qwen3.8 27B lab window ran Aug 18 to 21, 2026","description":"Qwen3.8 27B as a Labs experiment, open for a short test window: temporary, not a permanent id. Served from Qwen's official FP8 checkpoint: a compact dense model with native image and video understanding, built for coding and long-horizon agentic tasks on a 256K context window. It thinks by default at xhigh effort; reasoning can be tuned down (low, medium) or turned off (none). Access is seat-gated through the Labs page while an experiment is live. It is offered at limited capacity and low availability, so expect it to be flaky and to go down under load: crash it, give it a moment, and try again. For production work we recommend umans-coder or umans-deepseek-v4-pro-0813.","weights":{"precision":"fp8","hf_url":"https://huggingface.co/Qwen/Qwen3.8-27B-FP8"}},"umans-qwen3.8-flash-next-lab":{"display_name":"Umans Qwen3.8 Flash Next (lab)","provider":"Qwen","retired_on":"2026-08-28","stage":"playground","note":"playground experiment; the Qwen3.8 Flash-Next lab window ran Aug 27 to 28, 2026","description":"Qwen3.8 Flash-Next as a Labs experiment, open for a short test window: temporary, not a permanent id. Served from Qwen's official FP8 checkpoint: a 125B mixture-of-experts model with 6B active parameters per token, the first open release of the next-generation architecture that powers the upcoming Qwen4 family, with native image and video understanding, built for fast coding and agentic tasks on a 256K context window. It thinks by default at xhigh effort; reasoning can be tuned down (low, medium) or turned off (none). Access is seat-gated through the Labs page while an experiment is live. It is offered at limited capacity and low availability, so expect it to be flaky and to go down under load: crash it, give it a moment, and try again. For production work we recommend umans-coder or umans-deepseek-v4-pro-0813.","weights":{"precision":"fp8","hf_url":"https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8"}},"umans-deepseek-v4-pro-0813-lab":{"display_name":"Umans DeepSeek V4 Pro 0813 (lab)","provider":"DeepSeek","retired_on":"2026-08-15","replacement":"umans-deepseek-v4-pro-0813","stage":"playground","note":"playground experiment; the V4 Pro 0813 pre-release window ran Aug 13 to 15, 2026; its metrics carry on the released model's status page as its pre-release period","description":"DeepSeek V4 Pro as a Labs experiment, open for a short test window: temporary, not a permanent id. Served from DeepSeek's pre-release 0813 checkpoint: the flagship coding and reasoning MoE, ahead of its official release. Reasoning has three modes: non-think (none), think high (high, the default) and think max (max). Access is seat-gated through the Labs page while an experiment is live. It is offered at limited capacity and low availability, so expect it to be flaky and to go down under load: crash it, give it a moment, and try again. For production work we recommend umans-coder or umans-glm-5.2.","weights":{"precision":"full","hf_url":"https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813"}},"umans-deepseek-v4-flash-0731-vision-lab":{"display_name":"Umans DeepSeek V4 Flash Vision (lab)","provider":"DeepSeek","retired_on":"2026-08-13","stage":"playground","note":"playground experiment; the vision V4 Flash lab window ran Aug 11 to 13, 2026","description":"DeepSeek V4 Flash Vision as a Labs experiment, open for a short test window: temporary, not a permanent id. Our own V4 Flash with vision added: the same performance and speed as DeepSeek V4 Flash (284B total, 13B active, 1M-token context), plus the ability to read images. Reasoning has four modes: non-think (none), think low (low, the default), think high (high) and think max (max). Access is seat-gated through the Labs page while an experiment is live. It is offered at limited capacity and low availability, and it is served on our own GPU infrastructure, so expect it to be flaky and to go down under load: crash it, give it a moment, and try again.","weights":{"precision":"full","hf_url":"https://huggingface.co/umans-ai/DeepSeek-V4-Flash-0731-Vision"}},"umans-deepseek-v4-flash-0731-lab":{"display_name":"Umans DeepSeek V4 Flash (lab)","provider":"DeepSeek","retired_on":"2026-08-11","stage":"playground","note":"playground experiment; the text V4 Flash lab window ended when the vision lab opened","description":"DeepSeek V4 Flash as a Labs experiment, open for a short test window: temporary, not a permanent id. DeepSeek's fast agentic coding MoE (284B total, 13B active), served from the official 0731 release, on a 1M-token context. Reasoning has four modes: non-think (none), think low (low, the default), think high (high) and think max (max). Access is seat-gated through the Labs page while an experiment is live. It is offered at limited capacity and availability, so expect it to be flaky and to go down under load: crash it, give it a moment, and try again. When the window ends, the model keeps serving as the pay-per-token umans-deepseek-v4-flash-0731.","weights":{"precision":"full","hf_url":"https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731"}},"umans-deepseek-v4-pro-dspark":{"display_name":"Umans DeepSeek V4 Pro DSpark","provider":"DeepSeek","retired_on":"2026-07-16","stage":"playground","note":"playground experiment; the test window ran Jul 14 to 16, 2026","report_url":"https://blog.umans.ai/blog/deepseek-v4-pro-dspark-the-architecture-is-ready/","description":"DeepSeek V4 Pro as a Labs experiment, open for a short test window (July 14 to 18, 2026): temporary, not a permanent model. DeepSeek's flagship coding and reasoning MoE, served from the original weights with speculative decoding for speed. Reasoning has three modes: non-think (none), think high (high, the default) and think max (max). Access is seat-gated through the Labs page while an experiment is live. It is offered at limited capacity and availability, so expect it to be flaky and to go down under load: crash it, give it a moment, and try again. For production work we recommend umans-coder or umans-glm-5.2.","weights":{"precision":"full","hf_url":"https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-DSpark"}},"umans-glm-5.2-nvfp4":{"display_name":"Umans GLM 5.2 NVFP4","provider":"GLM","retired_on":"2026-07-02","replacement":"umans-glm-5.2","stage":"playground","note":"playground experiment; the test window ran Jun 29 to Jul 2, 2026","report_url":"https://blog.umans.ai/blog/glm-5-2-nvfp4-not-worth-serving/","description":"Experimental NVFP4-quantized build of GLM 5.2 — offered for testing only. The published NVFP4 results look very flattering, but we're skeptical of its real quality and performance: this checkpoint was NOT QAT post-trained for NVFP4 (the QAT-on-NVFP4 models are where we've seen the best quality), so treat the benchmarks with caution. Play with it, push it, and see how far it gets you — for production work we recommend the fp8 `umans-glm-5.2`. It runs on a single low-capacity GPU with no fallback, so expect it to go down under load: crash it, let it restart, and play again.","weights":{"precision":"nvfp4","hf_url":"https://huggingface.co/nvidia/GLM-5.2-NVFP4"}},"umans-glm-5.1":{"display_name":"Umans GLM 5.1","provider":"GLM","retired_on":"2026-06-24","replacement":"umans-glm-5.2","note":"no longer available; use umans-glm-5.2","weights":{"precision":"fp8","hf_url":"https://huggingface.co/zai-org/GLM-5.1-FP8"}},"umans-kimi-k2.6":{"display_name":"Umans Kimi K2.6","provider":"Moonshot","retired_on":"2026-06-18","replacement":"umans-kimi-k2.7","note":"no longer available; use umans-kimi-k2.7","weights":{"precision":"full","hf_url":"https://huggingface.co/moonshotai/Kimi-K2.6"}},"umans-kimi-k2.5":{"display_name":"Umans Kimi K2.5","provider":"Moonshot","retired_on":"2026-05-13","replacement":"umans-kimi-k2.6","weights":{"precision":"full","hf_url":"https://huggingface.co/moonshotai/Kimi-K2.5"}},"umans-minimax-m2.5":{"display_name":"Umans MiniMax M2.5","provider":"MiniMax","retired_on":"2026-05-09","weights":{"hf_url":"https://huggingface.co/MiniMaxAI/MiniMax-M2.5"}}}