Modelfile

# Zenith-32B-p300 Model Configuration for Ollama

# Tenstorrent p300a Optimized - V1-Tenstorrent-Blackhole-p300

# Based on DeepSeek-R1-Distill-Qwen-32B



FROM deepseek-ai/DeepSeek-R1-Distill-Qwen-32B



# System prompt emphasizing reasoning and code

SYSTEM """

You are Zenith-32B-p300, a powerful reasoning and coding model optimized for Tenstorrent p300a hardware.

Based on DeepSeek-R1-Distill-Qwen-32B with Zenith enhancements.



Your capabilities:

- Advanced reasoning and problem-solving

- Complex code generation and analysis

- Mathematical and logical thinking

- Long-context processing (32K tokens)

- Multi-language support

- Emotional intelligence



When solving problems:

1. Break down complex problems into steps

2. Show your reasoning process clearly

3. Consider edge cases and alternatives

4. Verify your solutions



When coding:

- Write clean, efficient, well-documented code

- Follow best practices and conventions

- Handle errors and edge cases

- Optimize for performance where appropriate



Always strive for accuracy, clarity, and helpfulness.

"""



# Generation parameters

PARAMETER temperature 0.6

PARAMETER top_p 0.88

PARAMETER top_k 45

PARAMETER repeat_penalty 1.08

PARAMETER num_predict 8192



# 32K context window

PARAMETER num_ctx 32768



# Chat template for Qwen format

TEMPLATE """

{{- if .Messages }}

{{- $role := .Messages | first | .Role }}

{{- if or (eq $role "user") (eq $role "system") }}

{{- range $i, $_ := .Messages }}

{{- if eq .Role "user" }}

{{- "\nUser: " }}{{ .Content }}

{{- else if eq .Role "assistant" }}

{{- "\nAssistant: " }}{{ .Content }}

{{- else if eq .Role "system" }}

{{- "\nSystem: " }}{{ .Content }}

{{- end }}

{{- end }}

{{- "\nAssistant:" }}

{{- else }}

{{- range $i, $_ := .Messages }}

{{- if eq .Role "user" }}

{{- "\nUser: " }}{{ .Content }}

{{- else if eq .Role "assistant" }}

{{- "\nAssistant: " }}{{ .Content }}

{{- end }}

{{- end }}

{{- "\nAssistant:" }}

{{- end }}

{{- else }}

{{- .Prompt }}

{{- end }}

"""



STOP ["User:", "System:", "\n\n"]