---
title: Qwen3.8-Smaug-Effithink-SLERP-27B
canonical_url: "https://www.modelscope.cn/models/Dalvlad/Qwen3.8-Smaug-Effithink-SLERP-27B"
md_url: "https://www.modelscope.cn/models/Dalvlad/Qwen3.8-Smaug-Effithink-SLERP-27B.md"
repository: Dalvlad/Qwen3.8-Smaug-Effithink-SLERP-27B
last_updated: 2026-10-10
license: apache-2.0
pipeline_tag: image-text-to-text
tasks:
  - image-text-to-text
model_type:
  - qwen3_5
architectures:
  - Qwen3_5ForConditionalGeneration
base_model:
  - abacusai/Smaug-Mini
  - Merkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2
base_model_relation: merge
parameters: 27.8B
tensor_type:
  - BF16
library_name:
  - safetensors
  - pytorch
frameworks:
  - pytorch
language:
  - en
  - zh
downloads: 20
stars: 0
tags:
  - qwen3.8
  - slerp
  - merge
  - coding
  - agentic
  - tool-use
  - long-context
  - vision
  - multimodal
  - mtp
---

# Qwen3.8-Smaug-Effithink-SLERP-27B

> Qwen3.8-Smaug-Effithink-SLERP-27B - Dalvlad 在 ModelScope 开源的模型。Qwen3.8-Smaug-Effithink-SLERP-27B

Dalvlad/Qwen3.8-Smaug-Effithink-SLERP-27B 是 ModelScope 魔搭社区上的 27.8B 参数image-text-to-text模型，采用 apache-2.0 许可，基于 abacusai/Smaug-Mini、Merkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 构建。

- **Repository**: Dalvlad/Qwen3.8-Smaug-Effithink-SLERP-27B
- **License**: apache-2.0
- **Tasks**: image-text-to-text
- **Parameters**: 27.8B
- **Base model**: abacusai/Smaug-Mini, Merkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2
- **Tags**: qwen3.8, slerp, merge, coding, agentic, tool-use, long-context, vision, multimodal, mtp
- **Downloads**: 20
- **Stars**: 0
- **Last updated**: 2026-10-10

Source: https://www.modelscope.cn/models/Dalvlad/Qwen3.8-Smaug-Effithink-SLERP-27B

---

# Qwen3.8-Smaug-Effithink-SLERP-27B

A **27B Qwen3.8 full non-skip SLERP merge** (t = 0.65 toward the coder):
the EfficientThink coding tower of
[Merkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2](https://modelscope.cn/models/Merkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2)
blended with
[abacusai/Smaug-Mini](https://modelscope.cn/models/abacusai/Smaug-Mini)
(Abacus.AI's agentic fine-tune) — both fine-tunes of the same
[Qwen/Qwen3.8-27B](https://modelscope.cn/models/Qwen/Qwen3.8-27B) backbone.
**The MTP draft head is merged too**, and the vision tower is byte-exact
Merkyor (so Merkyor's official mmproj GGUF works as-is).

GGUF quants live in the sibling repo
[Dalvlad/Qwen3.8-Smaug-Effithink-SLERP-27B-GGUF](https://modelscope.cn/models/Dalvlad/Qwen3.8-Smaug-Effithink-SLERP-27B-GGUF)
(Q4_K_M with MTP included + MTP draft + mmproj).

## The merge

- Both parents ship structurally identical checkpoints: 1199 tensors, same
  names/shapes, all BF16 (no bf16/f16 discrepancy).
- Weight-space diagnostics: parent cosine >= **0.9997 on every tensor**;
  lm_head rows **min/mean cosine 1.0000** (0 of 248,320 rows below 0.8) —
  no tensor needed anchoring and no lm_head exception was required. **Every
  non-vision tensor is interpolated** (850 text + 1 lm_head + 15 MTP) at
  t = 0.65 (65 % Coder390 / 35 % Smaug); the 333 vision tensors are
  byte-exact Merkyor.
- Verified post-merge: 360 independent SLERP spot-checks, all pass; vision
  byte-identity checked. Two-pass chunked spherical interpolation in
  float32, round-to-nearest-even BF16 output.

## Smoke tests (llama.cpp b393, Q4_K_M, CPU)

- **math** — 17*24 - 5 = **403**
- **factual** — capital of France -> Paris
- **datetime bug** (naive vs aware) — correct root cause + robust fix
- **hardest probe — CPython generator + `yield`-in-`finally` trace**
  (ground truth `1 / [1] / 3 / [1, 2] / caught done / [1, 2, 3]`, no
  `stop`): **all six lines exactly**
- **vision** — reads a true 3:30 clock face correctly

Playful merge, smoke-tested only — no benchmark has been run on the merged
model. For serious work use the parents.

## Weights

- `model-00000-of-00018.safetensors` ... `model-00018-of-00018.safetensors`
  (BF16, 64 hybrid layers = 48 linear-attention + 16 full-attention,
  hidden 5120, vocab 248320, native context 262144) + `mtp.*` draft-head
  tensors inside the shards + `model.visual.*` vision tower.

## Credits

- **[Qwen/Qwen3.8-27B](https://modelscope.cn/models/Qwen/Qwen3.8-27B)** —
  the shared backbone (Alibaba Cloud, Apache-2.0).
- **[abacusai/Smaug-Mini](https://modelscope.cn/models/abacusai/Smaug-Mini)**
  — agentic parent (Abacus.AI, Apache-2.0).
- **[Merkyor/Qwen3.8-27B-Coder390-EfficientThink-...](https://modelscope.cn/models/Merkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2)**
  — EfficientThink coding parent, vision tower + mmproj (Apache-2.0).

## Showcase: hard-trace reasoning (verified outcome)

The Heretic-Ara IQ4_XS-MTP GGUF built from this checkpoint was challenged
with a 7-trap CPython 3.13 semantics puzzle (closures, memoization,
`yield`-in-`finally` with `continue`, exception propagation, implicit
`del e`, `itertools.accumulate`, dict mutation) with an 8192-token
thinking budget: **7/7 output lines exact** against real CPython, all
mechanisms correctly explained. See the
[Heretic-Ara GGUF repo](https://modelscope.cn/models/Dalvlad/Qwen3.8-Smaug-Effithink-Heretic-Ara-16GB-VRAM-IQ4-XS-MTP-GGUF).