GEHIRNS/magi-1.0-2B

---

Introduction

magi-1.0-2B is the magi-1.0 mixed-precision recipe applied to Qwen3.5-2B, with no training: 1.36 GiB at a 32K context, capability density ρ = 1.071 against 0.903 for magi-1.0 (4B), measured under the same settings.

It is Qwen/Qwen3.5-2B quantized with magi-1.0's 13 tensor-class rules and the same quantizer, with an importance matrix computed on the 2B. Nothing else changes: the weights are the base model's, only stored at different precisions.

It scores 63.5%, 17.0 points below magi-1.0. Its main weakness is looping: under greedy decoding, 82% of its benchmark traces end in a repetition loop and run out the thinking budget (many still answer correctly after the cutoff). magi-1.1-2B is the same recipe plus a LoRA fine-tune aimed at these loops: it scores 66.1% (ρ = 1.149) at the same file size and loops on 77% of traces. Use magi-1.1-2B unless you want the training-free model.

Capability density

Capability density is ρ = −log₂(1 − score) / GiB: how many times the benchmark error rate is halved per gibibyte of memory. All three models here were measured the same way (llama.cpp, q4_0 KV cache, memory at a 32K context from llama-fit-params). The magi-1.0 card's chart used Ollama's memory figures, so its numbers are not comparable with these.

Benchmarks

Every column was run on one RTX 5070 (12 GB) with llama.cpp b11800 (llama-server) under one protocol:

• Tasks: 2,313, run with thinking on and a 4,096-token thinking budget (budget forcing), greedy decoding.
• Server: q4_0 KV cache with flash attention and micro-batch 128. Exception: magi-1.0's benchmark run used micro-batch 512. Its memory is shown at micro-batch 128, which passed a non-inferiority check on a 250-task sample.
• Code grading: by executing tests (EvalPlus for HumanEval+ and MBPP+). LiveCodeBench uses 150 problems from contests between 2025-01-18 and 2025-04-06.

Memory and ρ. Memory is the resident size llama-fit-params projects for a 32K context with one sequence: weights, KV cache, recurrent state and compute buffers, GPU plus host, without the CUDA context. These sizes are measured differently from the Ollama figures on the magi-1.0 card, so compare ρ only within this table.

Benchmark	magi-1.0 (4B) 2.61 GiB	magi-1.0-2B 1.36 GiB	magi-1.1-2B 1.36 GiB
Efficiency
Memory at 32K (GiB)	2.61	1.36	1.36
Benchmark score	80.49%	63.48%	66.06%
Capability density ρ	0.903	1.071	1.149
Decode speed (tok/s)	186	332	338
Traces ending in a loop	15.0%	82.4%	77.4%
Coding
HumanEval+	86.59	38.41	46.95
MBPP+	70.43	41.35	44.61
LiveCodeBench	24.00	0.67	1.33
Math
GSM8K	86.00	70.00	78.00
Science and Knowledge
ARC-Challenge	97.33	89.33	93.33
MMLU	79.00	61.00	64.00
OpenBookQA	93.50	74.00	77.00
SciQ	97.00	92.00	94.00
TruthfulQA	75.50	74.50	77.50
Commonsense
HellaSwag	73.50	54.50	50.50
WinoGrande	83.00	66.00	65.50
Tool Use
Function calling	100.00	100.00	100.00

Largest gaps to magi-1.0: competitive programming and code (LiveCodeBench, HumanEval+, MBPP+). magi-1.0-2B hits the thinking budget on 85% of tasks, mostly because it loops.

Decode speed: single-stream generation speed (llama-bench, 128 tokens, q4_0 KV), on the same GPU.

Recipe

The same rules as magi-1.0, applied to Qwen3.5-2B's 24 layers (18 Gated DeltaNet, 6 full attention):

Tensor group	Type	Share of parameters	bits/weight
FFN gate, up, down (all 24 layers)	IQ3_XXS	48.1%	3.06
Token embedding = tied output head	Q4_K	27.0%	4.50
DeltaNet input projection attn_qkv (18)	Q4_K	12.0%	4.50
Full-attention q, k, v, output (6 layers)	Q4_K	4.7%	4.50
Gated-attention output gate (18 DeltaNet layers)	IQ3_XXS	4.0%	3.06
DeltaNet output projection ssm_out (18)	Q4_K	4.0%	4.50
DeltaNet gates ssm_alpha, ssm_beta	Q8_0	<0.1%	8.50
Norms, convolution, A, Δt bias	F32	<0.1%	32.00

The quantizer is llama-quantize --pure from the PrismML-Eng/llama.cpp fork with magi-1.0's override file and base type IQ2_XXS. The importance matrix was computed on Qwen3.5-2B (BF16) from the same calibration text as magi-1.0: 32 reasoning traces for LiveCodeBench problems dated before the benchmark's contest window. File SHA-256: 94f4998fce4ecad4e7408d33dfe7a63814348feca1124be0aa21b7a15e439e9f.

How to run

llama.cpp (the evaluated configuration; uses the Qwen3.5 chat template embedded in the file):

llama-server -m magi-1.0-2B.gguf -c 32768 --jinja --flash-attn on --cache-type-k q4_0 --cache-type-v q4_0 -ub 128

Ollama:

ollama create magi-1.0-2B -f Modelfile
ollama run magi-1.0-2B

The benchmark used greedy decoding. For interactive use with thinking enabled, Qwen recommends temperature 0.6, top_p 0.95 and top_k 20, which also makes repetition loops much rarer.

Limitations

• Much weaker than magi-1.0 in absolute terms, most of all on code: LiveCodeBench 0.7%, against 24.0% for magi-1.0.
• Loops under greedy decoding: 82% of benchmark traces end in a repetition loop. Use sampling for interactive work, or magi-1.1-2B.
• Narrow evaluation: one runtime (llama.cpp b11800), one benchmark, one GPU, each configuration run once.
• The name is ours. "2B" follows Qwen's naming; the file holds 1,881,825,088 parameters. Text only: the base model's vision encoder is not included.

Related

• GEHIRNS/magi-1.0: the 4B model and the recipe this release reuses.
• GEHIRNS/magi-1.1-2B: this model plus a fine-tune that teaches it to finish its thinking.
• Paper on the recipe: Eason Sun, Where the Bits Go, Zenodo, 2026. DOI: 10.5281/zenodo.23630199.

License

The weights are Apache 2.0, inherited from Qwen/Qwen3.5-2B; the license text is included in this repository.