Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Studio: add MiniMax to Connections (#16011)
Nilayf8c1f2c92f62
diff --git a/studio/backend/core/inference/external_provider.py b/studio/backend/core/inference/external_provider.pyindex 4f2a7db67cf..99b0ab825e0 100644--- a/studio/backend/core/inference/external_provider.py+++ b/studio/backend/core/inference/external_provider.py@@ -786,6 +786,26 @@ def _mistral_thinking_spec(model: str) -> _MistralThinkingSpec:return _MistralThinkingSpec(models = (), style = "disabled")+_MINIMAX_EFFORTS = ("low", "medium", "high", "xhigh", "max")+++def _apply_minimax_reasoning_controls(+ body: dict[str, Any],+ model: str,+ enable_thinking: Optional[bool],+ reasoning_effort: Optional[str],+) -> None:+ """M3.1-Flash 400s on thinking off; M3 can disable it."""+ body["reasoning_split"] = True+ normalized = model.strip().lower()+ if normalized.startswith("minimax-m3.1"):+ effort = (reasoning_effort or "").strip().lower()+ if effort in _MINIMAX_EFFORTS:+ body["reasoning_effort"] = effort+ elif normalized == "minimax-m3" and enable_thinking is False:+ body["thinking"] = {"type": "disabled"}++def _apply_mistral_reasoning_controls(body: dict[str, Any],model: str,@@ -1762,8 +1782,8 @@ class ExternalProviderClient:if stream and self.provider_type in _USAGE_STREAM_OPTION_PROVIDERS:body["stream_options"] = {"include_usage": True}if max_tokens is not None:- # Newer OpenAI models (gpt-4o, gpt-5.x) reject max_tokens- if self.provider_type == "openai":+ # Newer OpenAI models reject max_tokens; MiniMax counts thinking against it.+ if self.provider_type in ("openai", "minimax"):body["max_completion_tokens"] = max_tokenselse:body["max_tokens"] = max_tokens@@ -1800,6 +1820,8 @@ class ExternalProviderClient:body["thinking"] = {"type": "disabled"}elif self.provider_type == "mistral":_apply_mistral_reasoning_controls(body, model, enable_thinking, reasoning_effort)+ elif self.provider_type == "minimax":+ _apply_minimax_reasoning_controls(body, model, enable_thinking, reasoning_effort)elif self.provider_type == "deepseek":_apply_deepseek_reasoning_controls(body, enable_thinking, reasoning_effort)elif self.provider_type in ("qwen", "qwencloud"):diff --git a/studio/backend/core/inference/external_tool_transport.py b/studio/backend/core/inference/external_tool_transport.pyindex b999c30b30f..c608befb703 100644--- a/studio/backend/core/inference/external_tool_transport.py+++ b/studio/backend/core/inference/external_tool_transport.py@@ -61,9 +61,11 @@ class OAICompatTransport:self._initial_message_count: int | None = Noneself._last_tools: list[dict[str, Any]] | None = None# Always, as the local GGUF loop does: preserve_thinking hides earlier turns, never this turn's tool rounds.- self.preserves_reasoning = client.provider_type in ("llama_cpp", "vllm") or _is_kimi_k3(- client.provider_type, model- )+ self.preserves_reasoning = client.provider_type in (+ "llama_cpp",+ "vllm",+ "minimax",+ ) or _is_kimi_k3(client.provider_type, model)def stream(self,diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.pyindex 44dd4d37a9c..62220729fe1 100644--- a/studio/backend/core/inference/providers.py+++ b/studio/backend/core/inference/providers.py@@ -286,6 +286,29 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {"notes": "Z.ai API key. China mainland: use base URL https://open.bigmodel.cn/api/paas/v4","body_omit": ("presence_penalty",),},+ "minimax": {+ "display_name": "MiniMax",+ "base_url": "https://api.minimax.io/v1",+ "default_models": [+ "MiniMax-M3.1-Flash",+ "MiniMax-M3",+ ],+ "model_capabilities": {+ "MiniMax-M3.1-Flash": {"vision": True},+ "MiniMax-M3": {"vision": True},+ },+ "supports_streaming": True,+ "supports_vision": True,+ "supports_tool_calling": True,+ "studio_tools": True,+ "auth_header": "Authorization",+ "auth_prefix": "Bearer ",+ "notes": "MiniMax API key. China: use base URL https://api.minimaxi.com/v1",+ # /v1/models lists every model on the account, retired M2.x (404) included; keep M3 and up.+ "model_id_allowlist": re.compile(r"^MiniMax-M(?:[3-9]|\d\d)", re.IGNORECASE),+ # Fixed at 1 / 0.95 server-side (https://platform.minimax.io/docs/api-reference/text-openai-api).+ "body_omit": ("temperature", "top_p", "presence_penalty"),+ },# Zen is pay as you go, Go a subscription; both translate every text model to /chat/completions."opencode": {"display_name": "OpenCode Zen",diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.pyindex acf81654ee1..604677769e2 100644--- a/studio/backend/routes/inference.py+++ b/studio/backend/routes/inference.py@@ -27608,6 +27608,8 @@ def _build_external_messages(# Thinking mode with tools answers 400 without it, and ignores it elsewhere# (api-docs.deepseek.com/guides/thinking_mode).or provider_type == "deepseek"+ # Tool turns need the thinking back (platform.minimax.io/docs/api-reference/text-openai-api).+ or provider_type == "minimax")result = []for msg in messages:diff --git a/studio/backend/tests/test_minimax_provider.py b/studio/backend/tests/test_minimax_provider.pynew file mode 100644index 00000000000..329f5a049b2--- /dev/null+++ b/studio/backend/tests/test_minimax_provider.py@@ -0,0 +1,102 @@+# SPDX-License-Identifier: AGPL-3.0-only+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0++import asyncio+import json++import httpx+import pytest++from core.inference import external_provider as ep+from core.inference.providers import get_base_url, get_provider_info, list_available_providers+++def _capture(monkeypatch, model, **kwargs) -> dict:+ captured: dict = {}++ def handler(request):+ captured["url"] = str(request.url)+ captured["body"] = json.loads(request.content)+ return httpx.Response(+ 200, headers = {"content-type": "text/event-stream"}, content = "data: [DONE]\n\n"+ )++ async def run():+ async with httpx.AsyncClient(transport = httpx.MockTransport(handler)) as http:+ monkeypatch.setattr(ep, "_client", lambda: http)+ client = ep.ExternalProviderClient("minimax", get_base_url("minimax"), "sk-test")+ async for _ in client.stream_chat_completion(+ messages = [{"role": "user", "content": "hi"}],+ model = model,+ temperature = 0.7,+ top_p = 0.95,+ max_tokens = 256,+ **kwargs,+ ):+ pass++ asyncio.run(run())+ return captured+++def test_minimax_listed():+ assert "minimax" in {p["provider_type"] for p in list_available_providers()}+++def test_minimax_lists_chat_models_live():+ info = get_provider_info("minimax")+ assert info.get("model_list_mode") != "curated"+ chat = ["MiniMax-M3.1-Flash", "MiniMax-M3"]+ retired = ["MiniMax-M2.7", "MiniMax-M2.5-highspeed", "MiniMax-M2"]+ listed = chat + retired + ["speech-2.8-hd", "MiniMax-Hailuo-2.3", "image-01"]+ assert [m for m in listed if info["model_id_allowlist"].match(m)] == chat+ assert info["default_models"] == chat+++def test_minimax_body_shape(monkeypatch):+ cap = _capture(monkeypatch, "MiniMax-M3")+ body = cap["body"]+ assert cap["url"] == "https://api.minimax.io/v1/chat/completions"+ assert body["max_completion_tokens"] == 256+ assert "max_tokens" not in body+ assert body["reasoning_split"] is True+ for field in ("temperature", "top_p", "presence_penalty", "thinking"):+ assert field not in body+++def test_minimax_m3_thinking_off(monkeypatch):+ cap = _capture(monkeypatch, "MiniMax-M3", enable_thinking = False)+ assert cap["body"]["thinking"] == {"type": "disabled"}++[email protected](("effort", "sent"), [("xhigh", "xhigh"), ("none", None)])+def test_minimax_m31_flash_effort_and_never_disabled(monkeypatch, effort, sent):+ cap = _capture(+ monkeypatch, "MiniMax-M3.1-Flash", enable_thinking = False, reasoning_effort = effort+ )+ assert cap["body"].get("reasoning_effort") == sent+ assert "thinking" not in cap["body"]+++def test_minimax_history_keeps_reasoning():+ from models.inference import ChatMessage+ from routes.inference import _build_external_messages++ calls = [+ {"id": "call_1", "type": "function", "function": {"name": "lookup", "arguments": "{}"}}+ ]+ history = [+ ChatMessage(role = "user", content = "hi"),+ ChatMessage(role = "assistant", content = "", reasoning_content = "plan", tool_calls = calls),+ ChatMessage(role = "tool", content = "42", tool_call_id = "call_1"),+ ChatMessage(role = "assistant", content = "answer", reasoning_content = "done"),+ ]+ out = _build_external_messages(history, True, provider_type = "minimax", model = "MiniMax-M3")+ replayed = [m.get("reasoning_content") for m in out if m["role"] == "assistant"]+ assert replayed == ["plan", "done"]+++def test_studio_tool_loop_replays_minimax_reasoning():+ from core.inference.external_tool_transport import OAICompatTransport+ client = ep.ExternalProviderClient("minimax", get_base_url("minimax"), "sk-test")+ assert OAICompatTransport(client, model = "MiniMax-M3").preserves_reasoning is Truediff --git a/studio/frontend/src/features/chat/external-providers.ts b/studio/frontend/src/features/chat/external-providers.tsindex bbc5282db77..0c037087118 100644--- a/studio/frontend/src/features/chat/external-providers.ts+++ b/studio/frontend/src/features/chat/external-providers.ts@@ -136,6 +136,7 @@ const VISION_CAPABLE_PROVIDER_TYPES = new Set<string>(["openai","anthropic","gemini",+ "minimax","openrouter",]);diff --git a/studio/frontend/src/features/chat/provider-capabilities.ts b/studio/frontend/src/features/chat/provider-capabilities.tsindex 0b4a634e99e..335ee5fbc6b 100644--- a/studio/frontend/src/features/chat/provider-capabilities.ts+++ b/studio/frontend/src/features/chat/provider-capabilities.ts@@ -655,6 +655,15 @@ const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {},qwen: OPENAI_COMPAT_BASE,zai: { ...OPENAI_COMPAT_BASE, presencePenalty: false },+ // Fixed server-side; the backend omits them too.+ minimax: {+ temperature: false,+ topP: false,+ topK: false,+ minP: false,+ repetitionPenalty: false,+ presencePenalty: false,+ },huggingface: OPENAI_COMPAT_BASE,// OpenRouter silently drops unsupported params, so surface every knob.openrouter: ALL_SUPPORTED,@@ -917,6 +926,20 @@ function withReasoningEffortStyle(caps: ReasoningCaps): ExternalReasoningCapabil};}+function resolveMiniMaxReasoningCapabilities(modelId: string): ExternalReasoningCapabilities {+ if (modelId.startsWith("minimax-m3.1")) {+ return withReasoningEffortStyle({+ supportsReasoning: true,+ supportsReasoningOff: false,+ reasoningEffortLevels: ["low", "medium", "high", "xhigh", "max"],+ });+ }+ if (modelId === "minimax-m3") {+ return withEnableThinkingStyle({ supportsReasoning: true, supportsReasoningOff: true });+ }+ return withEnableThinkingStyle();+}+function resolveKimiReasoningCapabilities(modelId: string): ExternalReasoningCapabilities {// Kimi has a boolean thinking toggle, not an effort scale: k2.6 on by default and// toggleable, k2-thinking always on, k2.5 and others none.@@ -1273,6 +1296,8 @@ export function getExternalReasoningCapabilities(? table: (catalogCapabilities("mistral", normalizedModel) ?? table);}+ case "minimax":+ return resolveMiniMaxReasoningCapabilities(modelForMatching);case "gemini": {// Compat gateways drop native thinkingConfig, so hide the ladder.if (isGeminiCustomOpenAICompatBase(options?.baseUrl)) {@@ -1346,7 +1371,10 @@ export function providerModelReplaysReasoning(providerType: string | null | undefined,modelId: string | null | undefined,): boolean {- return providerType === "kimi" && modelId?.trim().toLowerCase() === "kimi-k3";+ return (+ providerType === "minimax" ||+ (providerType === "kimi" && modelId?.trim().toLowerCase() === "kimi-k3")+ );}/** Sent as a llama-server / vLLM template kwarg, so the model's template must still read it. */diff --git a/studio/frontend/src/features/chat/provider-logo-path.ts b/studio/frontend/src/features/chat/provider-logo-path.tsindex c0b52222748..60872d08ff0 100644--- a/studio/frontend/src/features/chat/provider-logo-path.ts+++ b/studio/frontend/src/features/chat/provider-logo-path.ts@@ -13,6 +13,7 @@ const HUB_PROVIDER_IDS = new Map([["qwen", "qwen"],["qwencloud", "qwen"],["zai", "zai-org"],+ ["minimax", "minimax"],]);const HUB_LOGO_PATHS = new Map(diff --git a/studio/frontend/tests/external-preserve-thinking.test.ts b/studio/frontend/tests/external-preserve-thinking.test.tsindex c69a6e9f426..9ce41a350cc 100644--- a/studio/frontend/tests/external-preserve-thinking.test.ts+++ b/studio/frontend/tests/external-preserve-thinking.test.ts@@ -70,3 +70,12 @@ test("kimi-k3 replays reasoning whatever the preserve toggle says; k2.6 does not);}});++test("minimax replays reasoning whatever the preserve toggle says", () => {+ for (const enabled of [true, false]) {+ assert.deepEqual(+ evaluate(providerSupportsPreserveThinking, {providerType: "minimax"}, {preserveThinking: enabled}, true, "MiniMax-M3.1-Flash"),+ {replay: true, fields: {}},+ );+ }+});diff --git a/studio/frontend/tests/provider-logo-path.test.ts b/studio/frontend/tests/provider-logo-path.test.tsindex b1184e8f858..bf71eeafc2c 100644--- a/studio/frontend/tests/provider-logo-path.test.ts+++ b/studio/frontend/tests/provider-logo-path.test.ts@@ -21,6 +21,7 @@ test("connections use the model hub's canonical provider images", () => {["qwen", "qwen"],["qwencloud", "qwen"],["zai", "zai-org"],+ ["minimax", "minimax"],]) {const hubLogo = PROVIDER_LOGOS.find(({ id }) => id === hubId);assert.ok(hubLogo, hubId);@@ -40,7 +41,7 @@ test("every supported connection resolves to an existing public asset", () => {for (const type of ["openai", "openai_codex", "deepseek", "mistral", "huggingface", "kimi","qwen", "qwencloud", "gemini", "anthropic", "openrouter", "vllm", "ollama", "llama_cpp",- "liquid", "typesafe", "zai", "opencode", "opencode_go",+ "liquid", "typesafe", "zai", "minimax", "opencode", "opencode_go",]) {const path = providerLogoPath(type);assert.ok(path?.startsWith("/"), type);