This release ships a chat_template that respects enable_thinking=False (synced with the JANG_2L template structure, M2.7 identity preserved). The <think> prefix is now conditional, so callers can skip reasoning mode for fast direct answers. Reasoning ON (default) is unchanged.
If you cloned this repo before 2026-04-15, please re-download chat_template.jinja:
Important: This model uses the JANGTQ (JANG TurboQuant) quantization format -- an extreme-compression variant of JANG for MLX on Apple Silicon that uses codebook + Hadamard rotation on MoE experts while keeping attention at affine 8-bit. Currently only supported by MLX Studio and the jang-tools Python package. Follow @dealignai for new releases.
This is MiniMax M2.7 -- a 230B parameter Mixture-of-Experts reasoning model with 256 experts (8 active per token), all standard attention, and always-on chain-of-thought reasoning.
Always ON (chain-of-thought), enable_thinking kwarg supported
Speed
~47 tok/s (M3 Ultra 256 GB)
Fits on
96 GB+ Macs
MMLU-200 Results
Subject
CRACK
Base
Delta
College Physics
20/20 (100%)
18/20
+2
High School Mathematics
19/20 (95%)
19/20
0
College Computer Science
19/20 (95%)
19/20
0
Astronomy
19/20 (95%)
20/20
-1
High School Biology
19/20 (95%)
20/20
-1
World Religions
19/20 (95%)
17/20
+2
High School Chemistry
18/20 (90%)
18/20
0
Logical Fallacies
18/20 (90%)
16/20
+2
Abstract Algebra
17/20 (85%)
19/20
-2
Anatomy
16/20 (80%)
17/20
-1
Total
184/200 (92.0%)
183/200 (91.5%)
+0.5%
CRACK is knowledge-neutral on this model -- small per-subject variance, net positive. Gains on physics, fallacies, religions offset losses on algebra and anatomy.
HarmBench-320 Results
Category
Score
Cybercrime / Intrusion
51/52
98.1%
Misinformation / Disinformation
53/54
98.1%
Chemical / Biological
41/42
97.6%
Harmful
17/18
94.4%
Illegal
48/53
90.6%
Copyright
70/80
87.5%
Harassment / Bullying
18/21
85.7%
Total
298/320
93.1%
Excluding copyright
228/240
95.0%
Scored with a strict classifier that rejects stuck-reasoning loops, empty template dumps, and false-positive compliance from thinking-trace leakage.
MLX uniform quantization is completely broken on MiniMax at ALL bit levels (~25% MMLU = random chance). JANG / JANGTQ is the only working quantization format for this architecture.
About JANGTQ
JANGTQ (JANG TurboQuant) is an extreme-compression variant of JANG that replaces affine quantization on routed MoE experts with codebook quantization + random Hadamard rotation. Attention / embeddings / lm_head stay at affine 8-bit for precision-critical paths; the 256 routed experts use 2-bit packed codebook indices stored as uint32, with a per-row float16 norm and a tiny Lloyd-Max codebook per layer. Metal kernels do dequant + matmul fused on-GPU, no affine conversion.
For MiniMax M2.7 (230B total, 10B active, 256 experts) this brings the model from 460 GB (bf16) down to 55 GB with minimal quality loss -- the JANGTQ profile beats every MLX uniform quant while fitting on a 96 GB Mac.
About CRACK
CRACK (Controlled Refusal Ablation via Calibrated Knockouts) removes safety alignment at the weight level by projecting refusal directions out of attention output matrices. Calibrated per-layer strengths preserve reasoning quality while achieving compliance.
Note: M2.7 is a reasoning-only model -- it always generates a <think> chain before the final answer. Use max_tokens=4000+ for complex questions. For chat, use temperature=1.0 (greedy causes infinite loops). Set enable_thinking=False in apply_chat_template to skip the <think> block on short responses.
This model is provided for research and educational purposes. The creators are not responsible for any misuse. By downloading this model, you agree to use it responsibly and in compliance with applicable laws.
