Each message includes the full conversation history. So as we go:

  Turn 1:  [you msg]         → tokens for just turn 1
  Turn 2:  [turn 1 + you msg] → larger than turn 1
  Turn 3:  [turn 1 + 2 + you]   → even bigger  

Every new message re-sends everything you've said so far as context for the model to reason over. That's why a 20-message conversation costs way more than 20 separate 1-message conversations.

On the Ollama side, what you're seeing is probably a running meter that tallies up all the input+output tokens from your API calls. It counts both directions — what you send and what comes back.

Why it accelerates: Longer conversations → more context → more tokens per message → meter climbs faster as you go.

Caching: Ollama may cache shared context to reduce repeat transmission, but it's model/usage pattern dependent — not guaranteed.

Output length matters too — a short question from you + a long detailed answer from me = significant output tokens on my side of the meter.

no kidding! it sends the complete history every time!? Does that mean it's bettter to continue to open a new session per question or request? ~ @ajaxStardust

It's situational — depends on what you're doing:

Start a new session when:

• You're done with a topic and moving to something unrelated
• The prior context has no value for the new task
• You'd spend more tokens re-explaining everything than you'd save

Keep the session when:

• You're building on something (debugging code, iterative work)
• The prior context is genuinely useful to carry forward
• Re-establishing background would cost as much or more than continuing

The math: Say you have a 50-message conversation. Starting fresh saves ~49 messages of context... but if you need to re-explain 20 messages worth of background to get back up to speed, you've spent half your savings just getting there.

Best practice: It's not about fresh vs. continuing — it's about when to reset. For Ollama specifically, if you're done with a task, close it out and start fresh for the next one. But don't churn sessions mid-task just to save tokens.

Think of it like a document: you don't close a document every few paragraphs to "save memory." You close it when you're done with the chapter.

connected | idle agent main | session main (openclaw-tui) | ollama/minimax-m2.7:cloud | tokens 13k/197k (6%)

~ minimax-m2.7:cloud via OpenClaw on localhost

NVIDIA Nemotron 3 Ultra is a 550 billion parameter (55B active) open model from NVIDIA built for long-running, agentic workflows with fast and affordable performance across hundreds of tool calls.

Model highlights

• Built for long-running agents: Tuned for agent orchestration, coding agents, deep research, and complex enterprise workflows that run across hundreds of steps.
• 1M token context: Keep entire codebases, long tool histories, and research trails in context without losing the thread.
• Frontier reasoning, high efficiency: 550B total parameters with only 55B active per token, and optimized for NVFP4, NVIDIA’s 4-bit floating point format that packs the model into less memory and runs faster.

Benchmarks

Nemotron 3 Ultra leads on accuracy across agent productivity, instruction following, and long-context tasks, while delivering leading throughput—saving up to 30% on costs compared to other leading open models.

 Figure 1: Nemotron 3 Ultra leads among open models on agentic benchmarks for agent productivity, coding, and instruction following. 

Reference

•  NVIDIA Nemotron 3 Ultra blog 

Following the February release of the Qwen3.5 series, we’re pleased to share the first open-weight variant of Qwen3.6. Built on direct feedback from the community, Qwen3.6 prioritizes stability and real-world utility, offering developers a more intuitive, responsive, and genuinely productive coding experience.

Qwen3.6 Highlights

This release delivers substantial upgrades, particularly in

• Agentic Coding: the model now handles frontend workflows and repository-level reasoning with greater fluency and precision.

• Thinking Preservation: we’ve introduced a new option to retain reasoning context from historical messages, streamlining iterative development and reducing overhead.

qwen3.5:cloud - (untested) . jcyhsiao/qwen3.5cloud:latest - 0.0 GB (Cloud + Vision) - ⚡ [ agent claimed vision not supported ] Qwen by Alibaba. Qwen Studio offers comprehensive functionality spanning chatbot, image and video understanding, image generation, document processing, web search integration, tool utilization, and artifacts.

Ollama’s Cloud is officially licensed with MiniMax for commercial usage

In partnership with MiniMax, the M3 model on Ollama’s Cloud is US-based with zero data retention.

Highlights

• MiniMax M3 achieves top-tier performance on coding and agentic benchmarks, with autonomous task decomposition, tool invocation, and multi-step reasoning capabilities — providing a reliable foundation for AI coding assistants and automated workflows.

• Powered by the proprietary MiniMax Sparse Attention (MSA) architecture, M3 supports up to 1M tokens context window with a guaranteed minimum of 512K tokens. The 1M context is the infrastructure for long-range Agent tasks, long-range Coding, and long-video understanding.

• A natively multimodal model. The entire data pipeline was rebuilt to scale pretraining data to 100T+, with multimodal training from step zero achieving deep alignment between textual and visual semantic spaces. Multimodal is a native core capability, not a superficial add-on.

• On BrowseComp, M3 scores 83.5, surpassing Opus 4.7 (79.3), demonstrating strong autonomous browsing and information retrieval capabilities.

• Until now, only a handful of closed-source models could simultaneously achieve frontier coding capabilities, million-token context, and Multimodal. M3 is the first to bring complete frontier capability to the open world.

Benchmark

Architecture

MiniMax Sparse Attention (MSA) Architecture

The MSA architecture enables native ultra-long context pretraining. M3 supports up to 1M tokens context window with a guaranteed minimum of 512K tokens, delivering excellent inference latency and throughput at extreme context lengths. The 1M context is the infrastructure for long-range Agent tasks, long-range Coding, and long-video understanding.

• minimax-m2.7:cloud - 0.0 GB (Cloud) - ⚡⚡⚡⚡
• ollama launch hermes --model minimax-m3:cloud

 Gemma is a family of open models built by Google DeepMind. Gemma 4 models are multimodal, handling text and image input and generating text output.

Gemma 4 introduces key capability and architectural advancements :

• Reasoning – All models in the family are designed as highly capable reasoners, with configurable thinking modes.

• Extended Multimodalities – Processes Text, Image with variable aspect ratio and resolution support (all models)

• Diverse & Efficient Architectures – Offers Dense and Mixture-of-Experts (MoE) variants of different sizes for scalable deployment.

• Optimized for On-Device – Smaller models are specifically designed for efficient local execution on laptops and mobile devices.

• Increased Context Window – The small models feature a 128K context window, while the medium models support 256K.

• Enhanced Coding & Agentic Capabilities – Achieves notable improvements in coding benchmarks alongside native function-calling support, powering highly capable autonomous agents.

• Native System Prompt Support – Gemma 4 introduces native support for the system role, enabling more structured and controllable conversations.

Models

Ollama’s cloud

ollama run gemma4:31b-cloud

Edge models

The “E” in E2B and E4B stands for “effective” parameters, and are made for edge device deployments.

Effective 2B (E2B)

ollama run gemma4:e2b

Effective 4B (E4B)

ollama run gemma4:e4b

Workstation models

These models are designed for frontier intelligence locally.

12B

ollama run gemma4:12b

26B (Mixture of Experts model with 4B active parameters)

ollama run gemma4:26b

31B (Dense)

ollama run gemma4:31b

Benchmark Results

These models were evaluated against a large collection of different datasets and metrics to cover different aspects of text generation. Evaluation results marked in the table are for instruction-tuned models.

Model information

Mixture-of-Experts (MoE) Model

Best Practices

For the best performance, use these configurations and best practices:

1. Sampling Parameters

Use the following standardized sampling configuration across all use cases:

• temperature=1.0
• top_p=0.95
• top_k=64

2. Thinking Mode Configuration

Note that Ollama already handles the complexities of the chat template for you.

Compared to Gemma 3, the models use standard system , assistant , and user roles. To properly manage the thinking process, use the following control tokens:

• Trigger Thinking: Thinking is enabled by including the <|think|> token at the start of the system prompt. To disable thinking, remove the token.
• Standard Generation: When thinking is enabled, the model will output its internal reasoning followed by the final answer using this structure:
  <|channel>thought\n [Internal reasoning] <channel|>
• Disabled Thinking Behavior: For all models except for the E2B and E4B variants, if thinking is disabled, the model will still generate the tags but with an empty thought block:
  <|channel>thought\n<channel|> [Final answer]

3. Multi-Turn Conversations

• No Thinking Content in History : In multi-turn conversations, the historical model output should only include the final response. Thoughts from previous model turns must not be added before the next user turn begins.

4. Modality order

• For optimal performance with multimodal inputs, place image and/or audio content before the text in your prompt.

5. Variable Image Resolution

Aside from variable aspect ratios, Gemma 4 supports variable image resolution through a configurable visual token budget, which controls how many tokens are used to represent an image. A higher token budget preserves more visual detail

at the cost of additional compute, while a lower budget enables faster inference for tasks that don’t require fine-grained understanding.

• The supported token budgets are: 70 , 140 , 280 , 560 , and 1120 .
  
  • Use lower budgets for classification, captioning, or video understanding, where faster inference and processing many frames outweigh fine-grained detail.
  • Use higher budgets for tasks like OCR, document parsing, or reading small text.

Kimi K2.7 Code is a coding-focused agentic model built upon Kimi K2.6. With substantial improvements on real-world long-horizon coding tasks, it strengthens end-to-end task completion across complex software engineering workflows while improving token efficiency, reducing thinking-token usage by approximately 30% compared with Kimi K2.6.

Key Features

• Long-horizon coding : Substantial gains on realistic, end-to-end software engineering tasks across 10+ programming languages and a full production tech stack, spanning backend services, infrastructure, performance engineering, systems programming, security, frontend, and ML/data engineering.
• Improved token efficiency : Reduces thinking-token usage by approximately 30% compared with Kimi K2.6, while improving task completion on complex workflows.
• Stronger agentic tool use : Improved performance on multi-step tool calling and MCP-based environments, with interleaved thinking preserved across turns ( preserve_thinking ) for coherent multi-step coding sessions.
• Native multimodal : Supports image and video input via the MoonViT vision encoder, with a 256K token context window.
• kimi-k2.7-code:cloud
• kimi-k2.6:cloud - 0.0 GB (Cloud + Vision) - ⚡⚡⚡⚡ [ does well with finga.studio ]