Introducing Mach 2 Additive Medium
Qwen3.8-Flash-Next at 1.70 bits per weight, inferencing solely with addition rather than multiplication. 125.7 billion text parameters in 26.7 GB, retaining 96.3% of the BF16 model's scores across eight agentic and reasoning benchmarks.
The compressed weights fit on a single GPU with 32 GB or more, and run today through GGUF and the Mach fork of llama.cpp.
Frontier benchmarks, a tenth of the memory
Mach-2 Additive Medium against Qwen3.8-Flash-Next BF16, both run on the same harness, settings and tasks across knowledge, math, terminal, automation, coding and tool-use benchmarks.
Dot plot of 8 benchmark scores for Mach-2 Additive Medium beside its Qwen3.8-Flash-Next BF16 teacher. Humanity's Last Exam: 33.87 against 39.85, 85.0% retention. GPQA Diamond: 90.51 against 92.22, 98.1% retention. AIME 2026: 96.25 against 95.83, 100.4% retention. Terminal-Bench 2.1: 78.65 against 82.02, 95.9% retention. AutomationBench 1.0.6: 71.67 against 70.29, 102.0% retention. NL2Repo: 56.12 against 60.06, 93.4% retention. DeepSWE v1.1: 48.67 against 49.56, 98.2% retention. τ³-Banking: 45.1 against 46.41, 97.2% retention. Mean retention 96.3%.
Mach-2 Additive Medium
96.3%
Mean retention over 8 benchmarks
Weights on GPU
26.7 GB
251.4 GB at BF16
Bits per weight
1.70
Addition-only inference
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.Model card revision 2026-10-06
Scatter plot of average benchmark score against GB of weights on GPU. Mach-2 Additive Medium (1.70 bits · 26.7 GB · 96.3% of original) scores 58.2 at 26.7 GB. Qwen3.8 Flash-Next original (BF16 · 251.4 GB) scores 60.4 at 251.4 GB. Mach DeepSeek V4.1 Flash (2.0 bits · 146.8 GB · 98.7% of original) scores 63.1 at 146.8 GB. DeepSeek V4.1 Flash original (as released · 307.5 GB) scores 63.9 at 307.5 GB. Qwen3.8 27B original (BF16 · 54.0 GB) scores 51.5 at 54 GB. Bonsai 2 27B (PrismML) (1.76 bits · 5.9 GB · 68.4% of the 27B) scores 35.2 at 5.9 GB.
Average of HLE, Terminal-Bench 2.1, AutomationBench and DeepSWEFrom the model card's figure
Show the full table
| Benchmark | Tasks | Mach-2 Additive Medium | BF16 | Retention |
|---|---|---|---|---|
| Humanity's Last Exam | 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: Text subset. 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 1.0.6: 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 v1.1: 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.
Run it locally
Mach-2 Additive Medium runs from its compressed weights today. The 26.7 GB of weights fit on one GPU with 32 GB or more; the model's n-gram embedding table rides alongside at bf16 (102.4 GB).
- Weights
Run from the GGUF conversion with the Mach-1 fork of llama.cpp. The native repository holds the packed weights; decode.py documents the packed format.
- Sampling
temperature 1.0 · top_p 0.95 · top_k 20
Without these settings the model can loop.
- Speculative decoding
draft/mtp_q4g64.safetensors is the base model's MTP layer at 4 bits (group size 64), for speculative decoding with MLX.
- License
Apache-2.0. Base model Qwen/Qwen3.8-Flash-Next.