frozenlab/ qwen3.8-mtp

69 Downloads Updated 1 month ago

vision tools thinking 27b

ollama run frozenlab/qwen3.8-mtp:27b

curl http://localhost:11434/api/chat \
  -d '{
    "model": "frozenlab/qwen3.8-mtp:27b",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='frozenlab/qwen3.8-mtp:27b',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'frozenlab/qwen3.8-mtp:27b',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Applications

Claude Code

Claude Code ollama launch claude --model frozenlab/qwen3.8-mtp:27b

OpenCode

OpenCode ollama launch opencode --model frozenlab/qwen3.8-mtp:27b

Hermes Agent

Hermes Agent ollama launch hermes --model frozenlab/qwen3.8-mtp:27b

OpenClaw

OpenClaw ollama launch openclaw --model frozenlab/qwen3.8-mtp:27b

Models

Name

1 model

Size / Usage

Context

Input

qwen3.8-mtp:27b

18GB · 256K context window · Text, Image · 1 month ago

qwen3.8-mtp:27b

18GB

256K

Text, Image

Readme

qwen3.8-mtp

Qwen 3.8 27B with a shorter multi-token-prediction (MTP) draft for speculative decoding.

The weights, vision projector, chat template and license are byte-identical to the official qwen3.8:27b. Only the runtime parameters differ:

Parameter	`qwen3.8:27b`	`frozenlab/qwen3.8-mtp`
`draft_num_predict`	4	2
`min_p`	0	0.01

draft_num_predict sets how many tokens the MTP head drafts per step before the main model verifies them. A shorter draft is rejected less often, so less drafted work is thrown away when the text is hard to predict. min_p 0.01 cuts the long tail of very unlikely tokens.

If you already have qwen3.8:27b, pulling this model downloads no new weights.

Model

27.3B parameters, Q4_K_M
Context length: 262,144 tokens
Capabilities: completion, vision, tools, thinking
License: Apache 2.0

Requirements

Ollama 0.32.12 or newer.

Usage

ollama run frozenlab/qwen3.8-mtp:27b

curl http://localhost:11434/api/chat -d '{
  "model": "frozenlab/qwen3.8-mtp:27b",
  "messages": [{ "role": "user", "content": "Hello!" }]
}'

# qwen3.8-mtp

Qwen 3.8 27B with a shorter multi-token-prediction (MTP) draft for speculative decoding.

The weights, vision projector, chat template and license are byte-identical to the official
[`qwen3.8:27b`](https://ollama.com/library/qwen3.8). Only the runtime parameters differ:

| Parameter           | `qwen3.8:27b` | `frozenlab/qwen3.8-mtp` |
|---------------------|---------------|-------------------------|
| `draft_num_predict` | 4             | **2**                   |
| `min_p`             | 0             | **0.01**                |

`draft_num_predict` sets how many tokens the MTP head drafts per step before the main model
verifies them. A shorter draft is rejected less often, so less drafted work is thrown away
when the text is hard to predict. `min_p 0.01` cuts the long tail of very unlikely tokens.

If you already have `qwen3.8:27b`, pulling this model downloads no new weights.

## Model

- 27.3B parameters, Q4_K_M
- Context length: 262,144 tokens
- Capabilities: completion, vision, tools, thinking
- License: Apache 2.0

## Requirements

Ollama **0.32.12** or newer.

## Usage

```bash
ollama run frozenlab/qwen3.8-mtp:27b
```

```bash
curl http://localhost:11434/api/chat -d '{
  "model": "frozenlab/qwen3.8-mtp:27b",
  "messages": [{ "role": "user", "content": "Hello!" }]
}'
```

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)