File size: 3,921 Bytes
b0234bb 9536ece b0234bb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 | ---
license: apache-2.0
base_model: Qwen/Qwen3.8-Flash-Next
tags:
- qwen
- mach-2
- additive
- mixture-of-experts
---
# Mach-2 Additive Medium
Mach-2 Additive Medium is Qwen3.8-Flash-Next at **1.70 bits per weight**: 26.7 GB for the 125.7B text parameters, plus
the model's n-gram embedding table at bf16 (102.4 GB).
## Benchmarks
Retention is Mach-2 Additive Medium's score divided by that of Qwen3.8-Flash-Next BF16, with both run on the same
harness, settings and tasks.
| Benchmark | Tasks | Mach-2 Additive Medium | BF16 | Retention |
| --- | ---: | ---: | ---: | ---: |
| Humanity's Last Exam (text) | 2,158 | 33.87 | 39.85 | 85.0% |
| GPQA Diamond | 198 x 5 | 90.51 | 92.22 | 98.1% |
| AIME 2026 | 30 x 8 | 96.25 | 95.83 | 100.4% |
| Terminal-Bench 2.1 | 89 | 78.65 | 82.02 | 95.9% |
| AutomationBench 1.0.6 | 600 | 71.67 | 70.29 | 102.0% |
| NL2Repo | 98 | 56.12 | 60.06 | 93.4% |
| DeepSWE v1.1 | 113 | 48.67 | 49.56 | 98.2% |
| τ³-Banking | 97 x 5 | 45.10 | 46.41 | 97.2% |
- **Humanity's Last Exam**: up to 163,840 output tokens, graded by Artificial Analysis's judge.
- **GPQA Diamond**: mean accuracy over 5 samples per question. Answers that reached 81,920 tokens were re-asked at
163,840.
- **AIME 2026**: mean accuracy over 8 samples per problem, up to 81,920 output tokens.
- **Terminal-Bench 2.1**: Terminus 2, one attempt per task, 4 hours per task.
- **AutomationBench**: up to 50 steps per task, scored by Artificial Analysis's rule.
- **NL2Repo**: OpenHands without internet access; the 102 tasks BF16 finished, less 4 whose BF16 test run hit the
1-hour scoring limit. Test runs cut at that limit are left out rather than scored 0. Mach-2 Additive Medium's score
averages two runs.
- **DeepSWE**: Claude Code as the agent, all 113 tasks, up to 10.5 hours per task. A task counts as solved only if the
agent solved it within 700 steps.
- **τ³-Banking**: tau2-bench 1.0.1, 5 trials per task, GPT-5.4 mini as the simulated user and judge, as Artificial
Analysis runs it. Pass rate over the 459 conversations both models completed; 26 that outgrew the context window or
lost the simulated user to an API error are left out.
Empty answers and answers still unfinished at the token limit score 0. Agents and limits differ from those behind
Qwen's model card, so compare these columns with each other rather than with scores published elsewhere.
## Image input
`vision.safetensors` contains the original BF16 vision tower from the same pinned
Qwen3.8-Flash-Next checkpoint as the text weights. Its 333 `model.visual.*` tensors
are copied without quantization or numeric conversion; their source revision and
file checksum are recorded in `MANIFEST.json`.
Mach's image-input serving path loads this file directly from the downloaded model.
No separate base-model download is required for vision, including when serving offline.
The vision file is approximately 0.9 GB; the 1.70-bit figure above covers text
parameters only and excludes the vision tower and n-gram embedding table.
## Use
Run Mach-2 Additive Medium from its compressed weights with
[Mach-2 Additive Medium GGUF](https://huggingface.co/SyzygyResearch/Mach-2-Additive-Medium-GGUF) and the
[Mach-1 fork of llama.cpp](https://github.com/SyzygyResearch/llama.cpp-mach1). The 26.7 GB of weights fit on one GPU
with 32 GB or more.
Sample at temperature 1.0, top_p 0.95, top_k 20 (`generation_config.json`). Without these settings the model can loop.
This repository holds the packed weights; `decode.py` documents the packed format. `draft/mtp_q4g64.safetensors` is the
base model's MTP layer at 4 bits (group size 64) for speculative decoding with MLX.
|