Librechat - Revision history

Pelibby16: /* Local Models */

2026-05-05T15:51:15Z

Local Models

← Older revision		Revision as of 15:51, 5 May 2026
Line 115:		Line 115:
	'''Cons''': Falls behind Opus on the hardest agentic, search, and novel reasoning tasks.		'''Cons''': Falls behind Opus on the hardest agentic, search, and novel reasoning tasks.

	== Local Models ==		== Local Models (Ollama) ==
	=== Qwen 3.5 (qwen3.5:35b) ===		=== Qwen 3.5 (qwen3.5:35b) ===
	'''Best For''': Deep research, literature reviews, digesting massive textbooks, and synthesizing multiple PDFs.		'''Best For''': Deep research, literature reviews, digesting massive textbooks, and synthesizing multiple PDFs.

Pelibby16: /* Remote Models */

2026-04-07T19:58:31Z

Remote Models

← Older revision		Revision as of 19:58, 7 April 2026
Line 105:		Line 105:

	'''Cons''': Can be slower than Sonnet due to its size and depth of reasoning.		'''Cons''': Can be slower than Sonnet due to its size and depth of reasoning.

			=== Anthropic Claude Sonnet 4.6 ===
			'''Best for''': A strong balance of intelligence, speed, and cost. The best "daily driver" from Anthropic for most tasks.

			'''Target majors''': All majors. A great default online model for writing, coding, analysis, and general Q&A.

			'''Pros''': Very strong across the board while being faster and more affordable than Opus. Leads Opus in financial analysis (63.3%) and office tasks (GDPval-AA Elo: 1633). Strong coder (SWE-Bench Verified: 79.6%) and reasoner (GPQA Diamond: 89.9%).

			'''Cons''': Falls behind Opus on the hardest agentic, search, and novel reasoning tasks.

	== Local Models ==		== Local Models ==

Pelibby16: /* Remote Models */

2026-04-07T19:57:39Z

Remote Models

← Older revision		Revision as of 19:57, 7 April 2026
Line 96:		Line 96:

	'''Cons''': Benchmarks are lower than Pro across the board. Not ideal for tasks requiring peak reasoning or coding performance.		'''Cons''': Benchmarks are lower than Pro across the board. Not ideal for tasks requiring peak reasoning or coding performance.

			=== Anthropic Claude Opus 4.6 ===
			'''Best for''': The strongest Anthropic model. Excels at agentic tasks, novel problem-solving, complex tool use, and web browsing/search. Currently World Strongest Publicly available Model for Coding.

			'''Target majors''': Computer Science, Research, Business Analytics, any field requiring complex multi-step autonomous tasks.

			'''Pros''': Leads or ties for the top spot on many agentic benchmarks. It's the best model for novel problem-solving (ARC-AGI-2: 68.8%), agentic search (BrowseComp: 84.0%), and multidisciplinary reasoning (HLE with tools: 53.0%). It also leads in SWE-Bench Verified for agentic coding at 80.8% and graduate-level reasoning (GPQA Diamond: 91.3%).

			'''Cons''': Can be slower than Sonnet due to its size and depth of reasoning.

	== Local Models ==		== Local Models ==

Pelibby16: /* Remote Models */

2026-04-07T19:56:42Z

Remote Models

@@ Line 79: / Line 79: @@
 '''Cons''': As a cloud model, your prompts are sent to OpenAI's servers. Response times can vary based on server load.
-=== Test 2 ===
+=== Google Gemini 2.5 Pro ===
 == Local Models ==

Pelibby16: /* Remote Models */

2026-04-07T19:55:05Z

Remote Models

← Older revision		Revision as of 19:55, 7 April 2026
Line 70:		Line 70:
	= Models =		= Models =
	== Remote Models ==		== Remote Models ==
	=== ~~Test 1~~ ===		=== OpenAI ChatGPT 5.2 ===
			'''Best for''': The absolute cutting edge in reasoning, math, coding, and knowledge work. Also very good for writing.

			'''Target majors''': Law, Mathematics, Physics, Engineering, Computer Science, Graduate Researchers, anyone needing the highest possible accuracy.

			'''Pros''': State-of-the-art across nearly every benchmark. It achieves a perfect 100% on AIME 2025 (competition math) and leads in software engineering (80% SWE-Bench Verified). Its abstract reasoning (ARC-AGI-2: 52.9%) is in a class of its own compared to its predecessors.

			'''Cons''': As a cloud model, your prompts are sent to OpenAI's servers. Response times can vary based on server load.

	=== Test 2 ===		=== Test 2 ===

	== Local Models ==		== Local Models ==
	=== Qwen 3.5 (qwen3.5:35b) ===		=== Qwen 3.5 (qwen3.5:35b) ===

Pelibby16: /* Qwen3 Next Base (qwen3-next:80b) */

2026-04-07T19:51:37Z

Qwen3 Next Base (qwen3-next:80b)

← Older revision		Revision as of 19:51, 7 April 2026
Line 126:		Line 126:
	'''Cons''': Jack of all trades, master of none.		'''Cons''': Jack of all trades, master of none.

	'''~~Benchmark~~''': 76% GPQA Diamond, 61% IFBench		'''Benchmarks''': 76% GPQA Diamond, 61% IFBench

	= Tips for Getting Better Results =		= Tips for Getting Better Results =

Pelibby16: /* Qwen3 Next Base (qwen3-next:80b) */

2026-04-07T19:51:32Z

Qwen3 Next Base (qwen3-next:80b)

@@ Line 118: / Line 118: @@
 === Qwen3 Next Base (qwen3-next:80b) ===
 = Tips for Getting Better Results =
 * Be specific. The more context you give the AI, the better its response will be. Instead of "summarize this," try "summarize this in 3 bullet points for a non-technical audience."

Pelibby16: /* OpenAI GPT-OSS (gpt-oss:20b) */

2026-04-07T19:50:58Z

OpenAI GPT-OSS (gpt-oss:20b)

@@ Line 107: / Line 107: @@
 === OpenAI GPT-OSS (gpt-oss:20b) ===
 === Qwen3 Next Base (qwen3-next:80b) ===
 = Tips for Getting Better Results =

Pelibby16 at 19:50, 7 April 2026

2026-04-07T19:50:09Z

← Older revision		Revision as of 19:50, 7 April 2026
Line 69:		Line 69:

	= Models =		= Models =
	== Qwen 3.5 (qwen3.5:35b) ==		== Remote Models ==
			=== Test 1 ===
			=== Test 2 ===
			== Local Models ==
			=== Qwen 3.5 (qwen3.5:35b) ===
	'''Best For''': Deep research, literature reviews, digesting massive textbooks, and synthesizing multiple PDFs.		'''Best For''': Deep research, literature reviews, digesting massive textbooks, and synthesizing multiple PDFs.

Line 80:		Line 84:
	'''Benchmarks''': 70.2 IFBench (instruction following test), 84.2 GPQA Diamond (graduate level reasoning test), 85.2 MMMLU (multilingual knowledge), 89.3 OmniDocBench v1.5 (document recognition and understanding)		'''Benchmarks''': 70.2 IFBench (instruction following test), 84.2 GPQA Diamond (graduate level reasoning test), 85.2 MMMLU (multilingual knowledge), 89.3 OmniDocBench v1.5 (document recognition and understanding)

	== Qwen3 Coder Next (qwen3-coder-next:latest) ==		=== Qwen3 Coder Next (qwen3-coder-next:latest) ===
	'''Best For''': Complex programming assignments, navigating multi-file repositories, and long-form software engineering projects.		'''Best For''': Complex programming assignments, navigating multi-file repositories, and long-form software engineering projects.

Line 91:		Line 95:
	'''Benchmarks''': Achieves 70.6% on SWE-Bench Verified (a benchmark testing real-world GitHub issue resolution), matching or beating models that require 10x the computing power.		'''Benchmarks''': Achieves 70.6% on SWE-Bench Verified (a benchmark testing real-world GitHub issue resolution), matching or beating models that require 10x the computing power.

	== GLM-4.7 Flash (glm-4.7-flash:latest) ==		=== GLM-4.7 Flash (glm-4.7-flash:latest) ===
	'''Best For''': Web development, UI/UX prototyping, generating structured data, and multilingual translation.		'''Best For''': Web development, UI/UX prototyping, generating structured data, and multilingual translation.

Line 102:		Line 106:
	'''Benchmarks''': 59.2% SWE-Bench Verified (Lower than Qwen3 Coder Next)		'''Benchmarks''': 59.2% SWE-Bench Verified (Lower than Qwen3 Coder Next)

	== OpenAI GPT-OSS (gpt-oss:20b) ==		=== OpenAI GPT-OSS (gpt-oss:20b) ===
	== Qwen3 Next Base (qwen3-next:80b) ==		=== Qwen3 Next Base (qwen3-next:80b) ===
	= Tips for Getting Better Results =		= Tips for Getting Better Results =
	* Be specific. The more context you give the AI, the better its response will be. Instead of "summarize this," try "summarize this in 3 bullet points for a non-technical audience."		* Be specific. The more context you give the AI, the better its response will be. Instead of "summarize this," try "summarize this in 3 bullet points for a non-technical audience."

Pelibby16: /* Qwen3 Coder Next (qwen3-coder-next:latest) */

2026-04-07T19:43:54Z

Qwen3 Coder Next (qwen3-coder-next:latest)

← Older revision		Revision as of 19:43, 7 April 2026
Line 89:		Line 89:
	'''Cons''': It is a massive 80-billion-parameter model, taking up ~50GB of our server's disk space. It is highly specialized for code, meaning it will likely write dry, robotic essays if you try to use it for creative writing.		'''Cons''': It is a massive 80-billion-parameter model, taking up ~50GB of our server's disk space. It is highly specialized for code, meaning it will likely write dry, robotic essays if you try to use it for creative writing.

	'''~~Benchmark~~''': Achieves 70.6% on SWE-Bench Verified (a benchmark testing real-world GitHub issue resolution), matching or beating models that require 10x the computing power.		'''Benchmarks''': Achieves 70.6% on SWE-Bench Verified (a benchmark testing real-world GitHub issue resolution), matching or beating models that require 10x the computing power.

	== GLM-4.7 Flash (glm-4.7-flash:latest) ==		== GLM-4.7 Flash (glm-4.7-flash:latest) ==