rockmatrix-ai/EdgeRock-4B-FP8

EdgeRock-4B-FP8 — quantized release

Blockwise FP8 E4M3 weights with dynamic FP8 activations; a 5.592 GB weight package.

Run this version · GitHub · Original model

The overview, benchmark tables, and performance figures below describe the original EdgeRock-4B release. This package's quantization settings, launch commands, and conversion checks are in Quantized release.

---

EdgeRock

Strong Reasoning at Constant Memory
_{ROCK MATRIX}

English · 简体中文

EdgeRock brings strong reasoning to a fixed history budget. EdgeRock-4B, our first open-source release, supports text and image reasoning. Its constant-size history leaves more GPU memory available for sustained reasoning and concurrent serving. More model sizes are in preparation.

• Strong mathematical reasoning: 82.9% mean AIME pass@1, versus 18.8% for the strongest evaluated bounded-state baseline.
• O(1) history memory: fixed 72 MiB attention KV per sequence—99.1% less than Qwen3.5-4B.
• Higher serving throughput: 6.42× Qwen3.5-4B's output throughput at each model's best measured batch.

Quick start · Key results · How RMA works · Documentation

AIME is mean pass@1 across 2024–2026; KV is compared at 256K context. Serving throughput uses 32K input and 32K output on one Hopper 80GB GPU at each model's best measured batch. It counts output tokens only, with elapsed time including prefill.

Quick start

Model	Input modalities	Model weights
EdgeRock-4B	Text, images	Hugging Face · ModelScope

git clone https://github.com/rockmatrix-ai/EdgeRock.git
cd EdgeRock
python -m venv .venv
source .venv/bin/activate
python -m pip install ./wheels/transformers-5.7.0.dev0-py3-none-any.whl -e .
hf download rockmatrix-ai/EdgeRock --local-dir checkpoints/EdgeRock-4B \
  --include '*.safetensors' '*.json' '*.jinja' 'tokenizer*' 'vocab*' 'merges*' \
  --exclude 'evaluation_data/*'
cp config.json configuration_edgerock.py modeling_edgerock.py checkpoints/EdgeRock-4B/
python run.py demo --model checkpoints/EdgeRock-4B \
  --prompt 'Solve 17 × 23, explaining your reasoning.' --max-new-tokens 2048

For optimized inference and serving, see the inference guide.

Reasoning and multimodal capability

Strong reasoning at constant memory

EdgeRock-4B demonstrates strong mathematical reasoning at constant memory: 82.9% mean AIME 2024–2026 pass@1, versus 18.8% for RWKV-7 G1j in the 14B class. It leads 24 of 29 text benchmarks, with a 77.2% mean versus 53.4% for AHN-GDN Qwen2.5 14B.

All 17 bounded-state models

Bounded-state models. W2047 denotes Phi-3 Medium's configured 2,047-token attention window.

Model	29-benchmark mean	AIME mean	MATH-500	MMLU-Pro	LiveCodeBench v6	IFEval
EdgeRock-4B	77.2	82.9	97.5	79.8	61.6	88.0
AHN-GDN Qwen2.5 14B	53.4	11.4	80.0	63.5	32.9	77.4
RWKV-7 G1j 13.3B	52.3	18.8	77.2	47.6	30.7	77.8
Phi-3 Medium 4K Instruct (W2047)	45.4	2.2	49.0	54.0	19.1	61.4
RWKV-7 G1j 7.2B	43.5	8.1	64.2	46.2	21.3	66.5
RWKV-7 G1i 7.2B	42.9	9.4	68.6	44.9	18.8	53.4
AHN-GDN Qwen2.5 3B	40.9	4.2	66.2	42.8	14.6	58.6
RWKV-7 G1j 2.9B	37.0	2.2	48.2	37.9	12.1	56.4
RWKV-7 G1i 2.9B	35.3	1.8	43.8	37.8	11.6	43.1
QRWKV7 7B Instruct	32.6	0.0	41.0	23.0	12.4	59.0
Falcon3-Mamba 7B Instruct	31.6	1.2	41.8	26.1	9.3	64.7
RecurrentGemma 9B IT	29.9	0.4	22.8	13.4	7.9	39.7
RWKV-7 G1j 1.5B	28.4	1.1	31.4	16.6	6.3	46.2
RWKV-7 G1i 1.5B	27.9	0.3	28.2	13.3	7.2	37.0
ARWKV-R1 7B	18.1	0.1	33.2	8.8	2.6	10.4
xLSTM 7B	16.0	0.0	25.0	4.6	2.2	34.0
ARWKV-R1 1.5B	9.9	0.0	13.6	7.5	0.3	11.1

Comparison with compact models

EdgeRock-4B leads the compact-model comparison on MMLU-Pro and SuperGPQA and exceeds LFM2.5-2.6B on 28 of 29 text benchmarks, with constant persistent history. Baselines include Qwen3.5-4B, MiniCPM5-2B, and LFM2.5-2.6B, all using growing history.

Model	AIME mean	MMLU-Pro	SuperGPQA	GPQA-Diamond	LiveCodeBench v6	IFEval
EdgeRock-4B	82.9	79.8	60.0	64.6	61.6	88.0
Qwen3.5-4B	86.1	77.7	52.8	77.3	75.7	89.6
Gemma4-E4B-it	40.1	69.0	41.1	55.6	69.1	84.3
Nemotron-3-Nano-4B	57.1	56.8	37.4	43.4	66.4	89.5
Phi-4-mini-reasoning	38.9	38.5	32.2	48.0	35.3	45.7
MiniCPM5-2B	87.8	65.3	40.9	67.7	83.3	86.3
LFM2.5-2.6B	56.2	59.9	27.1	58.2	59.7	93.3

Backbone capability retention

Converting Qwen3.5-4B to EdgeRock-4B preserves 96.3% of its AIME mean, achieves 97.5% on MATH-500, and improves eight text benchmarks.

Full backbone-retention results

The figure ranks EdgeRock-4B minus Qwen3.5-4B in percentage points. Tables give absolute scores; retention is the ratio to Qwen3.5-4B (%).

Benchmark	EdgeRock-4B	Qwen3.5-4B	Retained (%)
AIME 2024	86.7	89.6	96.8
AIME 2025	77.4	80.8	95.8
AIME 2026	84.6	87.9	96.2
HMMT Feb 2025	71.8	74.6	96.2
HMMT Nov 2025	69.8	78.3	89.1
HMMT Feb 2026	63.6	65.9	96.5
MATH-500	97.5	98.0	99.5
AIME three-year mean	82.9	86.1	96.3

EdgeRock-4B vs Qwen3.5-4B: all 29 benchmarks

Benchmark	EdgeRock-4B	Qwen3.5-4B
GSM8K	94.5	94.8
MATH-500	97.5	98.0
AIME 2024	86.7	89.6
AIME 2025	77.4	80.8
AIME 2026	84.6	87.9
HMMT Feb 2026	63.6	65.9
MMLU	88.2	85.2
MMLU-Pro	79.8	77.7
MMLU-Redux	92.0	89.8
CMMLU	73.3	83.9
C-Eval	74.6	84.0
GPQA-Diamond	64.6	77.3
SuperGPQA	60.0	52.8
WinoGrande	77.9	87.1
HellaSwag	64.9	80.6
ARC-C	95.5	96.8
ARC-E	98.0	98.4
BBH	85.4	86.2
BBEH	28.9	34.6
HumanEval	96.7	86.0
MBPP	87.8	71.0
LiveCodeBench v6	61.6	75.7
IFEval	88.0	89.6
IFBench	65.3	60.3
DROP	77.2	76.0
MuSR	60.4	64.9
LogiQA2	72.2	82.2
HMMT Feb 2025	71.8	74.6
HMMT Nov 2025	69.8	78.3

Visual reasoning

EdgeRock-4B achieves 84.9% mean answer-level accuracy across six visual benchmarks, retaining 95.3% of Qwen3.5-4B's mean. Retention is the ratio to Qwen3.5-4B (%).

Six visual benchmarks and MME scores

Benchmark	EdgeRock-4B	Qwen3.5-4B	Retained (%)
MMBench EN	88.3	90.1	98.0
MMBench CN	89.0	91.4	97.4
POPE	85.4	88.8	96.2
MME	83.0	86.8	95.6
ChartQA	83.9	90.3	92.9
AI2D	80.0	87.0	92.0
Equal-weight mean	84.9	89.1	95.3

Standard MME scores (benchmark points):

MME metric	EdgeRock-4B	Qwen3.5-4B
Perception (points)	1,600.4	1,738.3
Cognition (points)	728.2	751.8
Total (points)	2,328.7	2,490.1

Inference efficiency

All runtime measurements use one Hopper 80GB GPU.

Constant history and GPU memory

EdgeRock-4B's attention-KV capacity stays at 72 MiB per sequence, versus Qwen3.5-4B's 8 GiB at 256K context—99.1% less with BF16 KV. Capacities below are calculated from model configurations.

Model	32K context (MiB)	64K context (MiB)	128K context (MiB)	256K context (MiB)	History growth
EdgeRock-4B	72	72	72	72	O(1)
Qwen3.5-4B	1,024	2,048	4,096	8,192	O(N)

History breakdown, GPU allocation, and runtime protocol

Batch-runtime persistent history: 97.76 MiB per sequence. The local ring stores 1,023 preceding positions; the decoder supplies the current token.

Component	MiB per sequence
Committed RMA anchors	32.00
Pending RMA candidates	8.00
Local-window KV	31.97
FP16 recurrent matrices	24.00
Convolution state	1.50
Metadata	0.29
Total persistent history	97.76

Measured GPU allocation, including weights, history, buffers, and workspace, peaks at 16.02 GiB for 128K input + 128K output at batch size 1 and 61.26 GiB across the 16 long-generation workloads at batch size 512 on one Hopper 80GB GPU.

Runtime protocol. 1K = 1,024 tokens. EdgeRock reports medians of three runs per workload; weighted rates pool decode tokens and time. Timing excludes loading, warmup, tokenization, and text decoding. Batch size counts submitted requests; vLLM schedules within its cache budget. Runtime details specify sampling, warmup, and workload construction.

EdgeRock-4B serves 512 concurrent long-generation sequences on one Hopper 80GB GPU. All runtime comparisons use BF16 weights/KV and ordinary autoregressive decoding with GPU-resident state; EdgeRock-4B uses FP32 GDN updates, with FP32 state storage at batch size 1 and FP16 storage for larger batches. Qwen3.5-4B uses FP32 GDN state.

The portable backend supports CPU and CUDA. For Hopper, the inference guide provides a fast profile with fused BF16 operations, CUDA Graphs, FP32 GDN updates, and FP32 state storage at batch size 1 or FP16 storage for larger batches.

Long-generation throughput

Best-batch output throughput

EdgeRock-4B delivers 6.42× Qwen3.5-4B's output throughput at 32K input + 32K output on one Hopper 80GB GPU: 12,124 tokens/s at batch size 512 versus 1,889 at batch size 64, each model's best measured batch.

All columns use 32K output. Cells give tokens/s: output tokens / total time including prefill, with batch size in parentheses.

Model	4K input	8K input	16K input	32K input
EdgeRock-4B	12,479 (512)	12,325 (512)	12,330 (512)	12,124 (512)
Qwen3.5-4B	3,103 (64)	2,853 (64)	2,472 (64)	1,889 (64)
Gemma 4 E4B	3,940 (128)	3,432 (128)	2,880 (128)	2,196 (64)
MiniCPM5-2B	3,152 (64)	2,938 (64)	2,575 (32)	2,117 (32)
LFM2.5-2.6B	1,673 (64)	1,771 (64)	1,695 (32)	1,727 (64)

Across 4K–32K inputs with 32K output, EdgeRock-4B's mean best-batch output throughput is 4.94× Qwen3.5-4B's and 7.18× LFM2.5-2.6B's.

Single-request decoding

EdgeRock-4B sustains 233.7 decode tokens/s, versus 151.3 for Qwen3.5-4B (1.54×), at 128K input + 128K output, batch size 1, on one Hopper 80GB GPU. Decode timing covers tokens after the first, including sampling and state updates; it excludes prefill.

Model	Decode tokens/s
EdgeRock-4B	233.7
Qwen3.5-4B	151.3
Gemma 4 E4B	133.8
MiniCPM5-2B	164.5
LFM2.5-2.6B	229.6

At 254K input + 2K output, EdgeRock-4B reaches 226.6 decode tokens/s versus Qwen3.5-4B's 133.9 (1.69×).

Concurrency scaling

EdgeRock-4B supports 512 concurrent long-generation sequences on one Hopper 80GB GPU, with 12,497 weighted decode tokens/s and 61.26 GiB peak allocation across 16 input/output length pairs.

Concurrency curves and complete batch sweep

At batch size 512 on one Hopper 80GB GPU, EdgeRock-4B reaches 12,497 weighted decode tokens/s, or 24.4 per sequence, across the 16 input/output pairs formed from 4K, 8K, 16K, and 32K lengths. Rates divide pooled subsequent output tokens by pooled decode time; memory values are workload peaks.

Batch size	Weighted decode tokens/s	Per-sequence tokens/s	Peak allocated GiB	Peak reserved GiB
32	4,128	129.0	18.94	19.49
64	6,685	104.5	20.16	21.80
128	9,049	70.7	26.74	27.92
256	11,231	43.9	39.90	41.77
320	11,074	34.6	46.49	50.69
384	12,099	31.5	53.06	57.27
512	12,497	24.4	61.26	66.97

Short-context decoding

At 4K input and 2K output, decode throughput scales from 229.6 tokens/s at batch size 1 to 12,481 at batch size 512, excluding prefill.

Short-context results

With 4K input + 2K output on one Hopper 80GB GPU, EdgeRock-4B scales from 229.6 decode tokens/s at batch size 1 to 12,481 at batch size 512, excluding prefill.

Batch size	Decode tokens/s	Per-sequence tokens/s	Peak allocated GiB
1	230	229.6	14.27
2	358	178.8	14.27
4	692	172.9	14.27
8	1,322	165.2	14.40
16	2,442	152.6	15.21
32	4,053	126.7	16.84
64	6,698	104.7	20.10
128	9,066	70.8	26.60
256	11,100	43.4	39.62
320	11,062	34.6	46.14
384	12,086	31.5	52.64
512	12,481	24.4	60.71

How Rock Memory Anchors work

A memory anchor is a learned key–value record retrievable after its source information leaves the local attention window. In EdgeRock's fixed-capacity bank, the writer creates records, admission selects which remain, and retrieval brings relevant records into the current computation.

EdgeRock-4B converts the eight full-attention layers of its Qwen3.5-4B backbone while retaining 24 Gated DeltaNet layers and the feed-forward blocks:

[GDN, GDN, GDN, sliding-window attention ∥ RMA] × 8

History component	Role	Capacity
Recurrent state	Summarize the evolving context	Fixed-dimensional matrices
Local attention	Resolve recent token interactions	1,024 positions
RMA anchor bank	Retain individually retrievable distant records	1,024 records per KV group
Pending candidates	Collect records for the next admission update	256 positions

Every 256 tokens, admission selects up to 1,024 records per KV group from existing anchors and new candidates. Queries read the committed bank; updates take effect in the next block. A query-dependent signed gate combines retrieved records with local attention.

Fixed component capacities give O(1) persistent history in sequence length and O(N) generation work for N output tokens at fixed model and memory dimensions.

Distant-memory retrieval

EdgeRock-4B achieves 94.1% distant-memory query accuracy, versus 0.4% with all eight RMA modules disabled, across 256 contexts of 16K–130,000 input tokens. All four queries are correct in 80.1% of contexts.

Memory evaluation and intervention results

Adaptation stages. Each stage uses the same 256 contexts and 1,024 queries. RLVR denotes reinforcement learning with verifiable rewards.

Adaptation stage	Query accuracy (%)	All four correct (%)
RMA adaptation (M0)	79.4	54.7
Joint adaptation (R)	96.1	87.9
Supervised fine-tuning (S)	93.8	82.4
Reinforcement learning (RLVR)	94.1	80.1

RMA contribution. Recurrence and local attention remain active in both settings.

Final-model setting	Correct queries / total	Query accuracy (%)	Complete contexts / total	All four correct (%)
RMA enabled	964 / 1,024	94.1	205 / 256	80.1
RMA disabled	4 / 1,024	0.4	0 / 256	0.0

Results by task. Each task contains 64 contexts and requires four correct answers per context.

Task	Complete contexts / total	All four correct (%)
Exact retrieval	58 / 64	90.6
State-update interpretation	56 / 64	87.5
Cross-document combination	57 / 64	89.1
Variable binding	34 / 64	53.1
All tasks	205 / 256	80.1

Paired evidence changes. Success requires both answers in a pair to be correct.

Change to distant evidence	Correct pairs / total	Pair accuracy (%)
Answer should change	230 / 256	89.8
Answer should stay the same	247 / 256	96.5

Full benchmark results

EdgeRock-4B is compared with 16 bounded-state baselines and 11 growing-history models across all 29 text benchmarks. Scores use each benchmark's primary metric on a 0–100 scale; bold marks column maxima, including ties. AIME reports mean pass@1, and DROP reports F1.

Bounded-state comparison: all 29 benchmarks

Phi-3 Medium uses a configured 2,047-token attention window.

Mathematics: bounded-state models

Model	GSM8K	MATH-500	AIME 2024	AIME 2025	AIME 2026	HMMT Feb 2025	HMMT Nov 2025	HMMT Feb 2026
EdgeRock-4B	94.5	97.5	86.7	77.4	84.6	71.8	69.8	63.6
ARWKV-R1 1.5B	24.8	13.6	0.0	0.0	0.0	0.4	0.8	1.1
RWKV-7 G1i 1.5B	60.3	28.2	0.0	0.4	0.4	1.7	5.0	0.0
RWKV-7 G1j 1.5B	57.2	31.4	1.2	2.1	0.0	1.7	4.2	0.0
RWKV-7 G1i 2.9B	71.7	43.8	1.7	1.7	2.1	1.7	5.0	0.8
RWKV-7 G1j 2.9B	74.1	48.2	0.4	4.6	1.7	2.1	5.0	0.4
AHN-GDN Qwen2.5 3B	83.8	66.2	7.1	2.1	3.3	2.9	6.7	4.9
ARWKV-R1 7B	51.2	33.2	0.0	0.0	0.4	0.8	1.7	1.1
Falcon3-Mamba 7B Instruct	61.4	41.8	1.7	0.8	1.2	0.8	5.4	2.7
QRWKV7 7B Instruct	65.5	41.0	0.0	0.0	0.0	0.0	6.7	0.0
xLSTM 7B	43.4	25.0	0.0	0.0	0.0	1.7	5.0	1.5
RWKV-7 G1i 7.2B	85.1	68.6	8.8	11.7	7.9	2.1	5.0	3.4
RWKV-7 G1j 7.2B	81.8	64.2	8.3	11.2	4.6	4.2	4.2	0.8
RecurrentGemma 9B IT	51.7	22.8	0.4	0.8	0.0	1.2	5.4	1.9
RWKV-7 G1j 13.3B	94.6	77.2	17.5	27.5	11.2	3.3	5.8	12.1
AHN-GDN Qwen2.5 14B	93.6	80.0	13.8	10.0	10.4	7.5	10.0	4.9
Phi-3 Medium 4K Instruct	85.3	49.0	3.3	3.3	0.0	6.7	2.5	3.0

Knowledge and science: bounded-state models

Model	MMLU	MMLU-Pro	MMLU-Redux	CMMLU	C-Eval	GPQA-Diamond	SuperGPQA
EdgeRock-4B	88.2	79.8	92.0	73.3	74.6	64.6	60.0
ARWKV-R1 1.5B	17.4	7.5	17.9	15.6	15.2	13.1	5.8
RWKV-7 G1i 1.5B	49.7	13.3	55.4	33.6	33.0	26.8	17.7
RWKV-7 G1j 1.5B	49.3	16.6	53.8	35.3	33.3	26.3	16.1
RWKV-7 G1i 2.9B	63.6	37.8	63.4	36.1	32.5	26.8	21.6
RWKV-7 G1j 2.9B	64.7	37.9	58.9	37.9	36.5	31.8	21.8
AHN-GDN Qwen2.5 3B	62.7	42.8	65.7	62.4	62.9	31.3	23.0
ARWKV-R1 7B	32.3	8.8	35.2	29.0	29.0	13.1	8.8
Falcon3-Mamba 7B Instruct	57.7	26.1	64.8	33.4	35.2	20.7	12.2
QRWKV7 7B Instruct	64.6	23.0	34.1	57.7	56.0	8.6	15.4
xLSTM 7B	19.6	4.6	37.9	16.9	19.5	12.1	5.3
RWKV-7 G1i 7.2B	69.4	44.9	73.5	49.8	51.6	30.3	25.7
RWKV-7 G1j 7.2B	67.5	46.2	72.7	50.8	48.9	36.9	25.3
RecurrentGemma 9B IT	55.5	13.4	55.0	43.4	41.7	24.7	19.5
RWKV-7 G1j 13.3B	76.7	47.6	81.8	67.4	66.3	42.4	31.8
AHN-GDN Qwen2.5 14B	77.2	63.5	83.5	75.0	74.2	40.9	35.5
Phi-3 Medium 4K Instruct	76.3	54.0	81.4	53.3	51.3	26.8	25.6

Commonsense and reasoning: bounded-state models

Model	WinoGrande	HellaSwag	ARC-C	ARC-E	BBH	BBEH
EdgeRock-4B	77.9	64.9	95.5	98.0	85.4	28.9
ARWKV-R1 1.5B	31.4	7.4	19.6	22.3	9.8	1.6
RWKV-7 G1i 1.5B	46.6	33.4	68.6	83.0	27.8	4.6
RWKV-7 G1j 1.5B	49.3	24.6	66.2	80.8	31.4	5.0
RWKV-7 G1i 2.9B	57.4	62.7	81.9	92.2	30.3	4.5
RWKV-7 G1j 2.9B	58.1	65.9	83.3	91.6	39.5	6.5
AHN-GDN Qwen2.5 3B	61.7	52.3	76.6	86.1	39.5	4.6
ARWKV-R1 7B	34.4	10.0	47.4	56.8	16.6	2.4
Falcon3-Mamba 7B Instruct	57.5	38.6	74.8	87.6	32.8	3.4
QRWKV7 7B Instruct	61.8	48.0	86.3	95.4	37.4	1.9
xLSTM 7B	28.3	7.9	25.1	28.2	12.8	3.3
RWKV-7 G1i 7.2B	60.4	70.2	84.5	94.5	38.8	7.0
RWKV-7 G1j 7.2B	59.4	71.7	83.7	94.5	41.4	7.7
RecurrentGemma 9B IT	56.4	45.8	73.3	87.2	35.5	3.2
RWKV-7 G1j 13.3B	72.1	77.4	91.8	96.3	48.7	7.7
AHN-GDN Qwen2.5 14B	74.0	67.9	91.5	94.0	52.0	5.3
Phi-3 Medium 4K Instruct	79.6	74.0	89.8	95.3	53.3	3.1

Code and instruction following: bounded-state models

Model	HumanEval	MBPP	LiveCodeBench v6	IFEval	IFBench
EdgeRock-4B	96.7	87.8	61.6	88.0	65.3
ARWKV-R1 1.5B	4.3	5.4	0.3	11.1	8.3
RWKV-7 G1i 1.5B	42.1	36.6	7.2	37.0	14.7
RWKV-7 G1j 1.5B	41.5	38.6	6.3	46.2	16.0
RWKV-7 G1i 2.9B	58.5	49.2	11.6	43.1	17.0
RWKV-7 G1j 2.9B	58.5	48.2	12.1	56.4	17.3
AHN-GDN Qwen2.5 3B	67.7	49.0	14.6	58.6	26.3
ARWKV-R1 7B	16.5	21.2	2.6	10.4	8.3
Falcon3-Mamba 7B Instruct	34.1	39.2	9.3	64.7	18.7
QRWKV7 7B Instruct	62.2	30.2	12.4	59.0	21.3
xLSTM 7B	36.6	9.2	2.2	34.0	14.7
RWKV-7 G1i 7.2B	65.2	59.0	18.8	53.4	17.7
RWKV-7 G1j 7.2B	72.6	58.0	21.3	66.5	18.3
RecurrentGemma 9B IT	37.2	19.0	7.9	39.7	12.7
RWKV-7 G1j 13.3B	86.0	67.4	30.7	77.8	27.3
AHN-GDN Qwen2.5 14B	75.6	68.6	32.9	77.4	33.3
Phi-3 Medium 4K Instruct	73.8	59.4	19.1	61.4	22.0

Reading and multi-step reasoning: bounded-state models

Model	DROP F1	MuSR	LogiQA2
EdgeRock-4B	77.2	60.4	72.2
ARWKV-R1 1.5B	5.6	8.6	17.5
RWKV-7 G1i 1.5B	23.0	19.3	38.7
RWKV-7 G1j 1.5B	24.9	29.1	36.1
RWKV-7 G1i 2.9B	24.6	31.9	47.6
RWKV-7 G1j 2.9B	30.7	34.5	45.4
AHN-GDN Qwen2.5 3B	32.8	37.8	50.1
ARWKV-R1 7B	11.1	15.4	26.8
Falcon3-Mamba 7B Instruct	19.9	29.1	41.4
QRWKV7 7B Instruct	21.2	2.6	33.2
xLSTM 7B	29.0	19.3	20.0
RWKV-7 G1i 7.2B	33.3	49.6	53.1
RWKV-7 G1j 7.2B	34.2	48.6	56.4
RecurrentGemma 9B IT	37.1	35.2	40.2
RWKV-7 G1j 13.3B	51.5	51.4	66.6
AHN-GDN Qwen2.5 14B	66.5	56.4	72.3
Phi-3 Medium 4K Instruct	51.5	51.1	62.7

Comparison with growing-history models: all 29 benchmarks

EdgeRock-4B retains constant persistent history; the eleven comparison models retain history that grows with sequence length.

Mathematics: EdgeRock vs growing-history models

Model	GSM8K	MATH-500	AIME 2024	AIME 2025	AIME 2026	HMMT Feb 2025	HMMT Nov 2025	HMMT Feb 2026
EdgeRock-4B	94.5	97.5	86.7	77.4	84.6	71.8	69.8	63.6
Qwen3.5-4B	94.8	98.0	89.6	80.8	87.9	74.6	78.3	65.9
Gemma4-E2B-it	90.1	78.8	38.3	33.8	33.8	21.7	17.5	20.1
MiniCPM5-2B	93.6	97.0	87.5	85.4	90.4	76.7	74.6	63.3
LFM2.5-2.6B	91.0	92.7	63.4	56.3	48.9	34.7	42.9	32.9
Qwen3.5-2B	87.6	86.8	41.2	34.2	35.4	26.2	23.3	25.4
Granite4.2-3B	93.8	97.8	85.0	78.8	82.9	75.8	72.5	59.5
Phi-4-mini-reasoning	93.6	91.6	45.4	32.5	38.8	25.0	28.7	24.6
Gemma4-E4B-it	91.3	88.2	42.1	40.0	38.3	25.4	32.9	28.8
Nemotron-3-Nano-4B	92.4	94.8	60.8	56.7	53.8	40.0	47.1	36.4
Granite4.2-8B	94.4	98.0	88.3	87.9	91.7	81.7	86.7	73.9
MiniCPM4.1-8B	93.6	96.4	77.1	70.0	70.8	58.3	59.6	49.2

Knowledge and science: EdgeRock vs growing-history models

Model	MMLU	MMLU-Pro	MMLU-Redux	CMMLU	C-Eval	GPQA-Diamond	SuperGPQA
EdgeRock-4B	88.2	79.8	92.0	73.3	74.6	64.6	60.0
Qwen3.5-4B	85.2	77.7	89.8	83.9	84.0	77.3	52.8
Gemma4-E2B-it	72.2	47.0	78.2	60.2	61.2	46.5	34.1
MiniCPM5-2B	79.6	65.3	85.3	79.2	78.3	67.7	40.9
LFM2.5-2.6B	75.3	59.9	81.7	57.2	59.5	58.2	27.1
Qwen3.5-2B	72.5	62.8	79.5	68.2	68.7	54.5	39.3
Granite4.2-3B	77.2	66.4	81.6	59.1	60.8	58.6	39.5
Phi-4-mini-reasoning	77.5	38.5	78.3	45.3	42.4	48.0	32.2
Gemma4-E4B-it	81.2	69.0	86.4	68.7	68.6	55.6	41.1
Nemotron-3-Nano-4B	75.8	56.8	81.5	54.6	56.4	43.4	37.4
Granite4.2-8B	78.6	72.0	86.3	55.1	58.4	61.6	45.7
MiniCPM4.1-8B	82.7	67.0	87.8	80.2	79.0	51.5	40.6

Commonsense and reasoning: EdgeRock vs growing-history models

Model	WinoGrande	HellaSwag	ARC-C	ARC-E	BBH	BBEH
EdgeRock-4B	77.9	64.9	95.5	98.0	85.4	28.9
Qwen3.5-4B	87.1	80.6	96.8	98.4	86.2	34.6
Gemma4-E2B-it	63.2	50.9	89.2	94.8	80.3	11.6
MiniCPM5-2B	69.8	52.2	91.9	97.0	82.1	25.4
LFM2.5-2.6B	67.9	40.7	91.6	96.5	76.7	13.1
Qwen3.5-2B	71.5	59.3	91.4	95.5	79.1	14.9
Granite4.2-3B	74.0	67.4	92.4	97.1	85.5	24.5
Phi-4-mini-reasoning	75.8	62.6	90.2	94.4	72.5	12.9
Gemma4-E4B-it	76.5	53.5	94.3	97.7	84.1	17.4
Nemotron-3-Nano-4B	65.6	46.3	91.9	96.7	79.3	10.8
Granite4.2-8B	76.2	63.1	90.9	94.6	88.9	31.5
MiniCPM4.1-8B	78.2	49.8	93.4	97.4	86.6	14.0

Code and instruction following: EdgeRock vs growing-history models

Model	HumanEval	MBPP	LiveCodeBench v6	IFEval	IFBench
EdgeRock-4B	96.7	87.8	61.6	88.0	65.3
Qwen3.5-4B	86.0	71.0	75.7	89.6	60.3
Gemma4-E2B-it	90.9	79.6	59.6	82.4	29.0
MiniCPM5-2B	97.6	94.8	83.3	86.3	64.0
LFM2.5-2.6B	92.9	73.6	59.7	93.3	56.2
Qwen3.5-2B	54.9	35.4	25.0	81.5	43.7
Granite4.2-3B	97.0	92.6	75.9	92.2	74.3
Phi-4-mini-reasoning	79.9	64.0	35.3	45.7	14.3
Gemma4-E4B-it	93.9	79.2	69.1	84.3	34.7
Nemotron-3-Nano-4B	90.9	72.2	66.4	89.5	57.7
Granite4.2-8B	96.3	95.4	85.6	94.6	77.0
MiniCPM4.1-8B	98.2	92.2	70.4	73.4	23.7

Reading and multi-step reasoning: EdgeRock vs growing-history models

Model	DROP F1	MuSR	LogiQA2
EdgeRock-4B	77.2	60.4	72.2
Qwen3.5-4B	76.0	64.9	82.2
Gemma4-E2B-it	72.9	58.9	59.4
MiniCPM5-2B	76.0	65.3	73.7
LFM2.5-2.6B	72.3	60.4	63.2
Qwen3.5-2B	73.7	61.3	66.3
Granite4.2-3B	83.7	65.1	69.5
Phi-4-mini-reasoning	50.6	56.7	58.5
Gemma4-E4B-it	72.9	62.2	72.6
Nemotron-3-Nano-4B	72.6	58.2	67.9
Granite4.2-8B	76.9	62.4	73.8
MiniCPM4.1-8B	78.3	64.5	75.7

Capability evaluation protocol

The 29-benchmark mean weights tasks equally after their subject/subtask aggregation. AIME pools responses within each edition, then averages editions equally. EdgeRock text evaluation uses temperature 0.6, top-p 0.95, top-k 20, and a 131,072-token response cap. Repetition counts vary by model and benchmark; evaluation details and Appendix A of the report give sampling and scoring procedures.

Documentation

• Inference guide · Runtime and reproduction · Evaluation details
• Technical report · LaTeX source · Citation metadata
• Runtime source · Documentation index for AI readers

Frequently asked questions

What is EdgeRock-4B?

EdgeRock-4B is Rock Matrix's first open-source 4B multimodal reasoning model. It supports text and image reasoning with constant-size persistent history. Weights and inference code are available.

What stays constant in O(1) memory?

In the batch runtime, EdgeRock-4B's persistent history stays at 97.76 MiB per sequence, including attention KV and recurrent state, as sequence length grows. Total GPU memory also includes weights and execution buffers and grows with concurrency.

What do memory anchors add to sliding-window attention?

Sliding-window attention reads recent tokens. Rock Memory Anchors preserve selected distant information as individually retrievable key–value records in a fixed-capacity bank, extending direct retrieval beyond the local window.

Does EdgeRock require quantization or speculative decoding?

EdgeRock-4B's reported runtime uses unquantized BF16 weights and KV, FP32 GDN updates, and ordinary autoregressive decoding. GDN state storage is FP32 at batch size 1 and FP16 for larger batches. Fixed-capacity history provides the memory savings; CUDA Graphs and fused operators accelerate execution. See the runtime configuration.

Citation

@techreport{edgerock2026,
  title = {EdgeRock: Strong Reasoning at Constant Memory},
  author = {Ji, Huishan and Zhang, Weiming and Ye, Qirui and Xu, Jinchao and Guo, Yike and Xu, Hua},
  institution = {Rock Matrix},
  year = {2026},
  eprint = {2610.10118},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.10118}
}

License

Runtime code is distributed under Apache-2.0. Model weights are governed by the model license. See NOTICE for attribution and third-party terms.

---

Quantized release: EdgeRock-4B-FP8

EdgeRock's constant-memory reasoning architecture in a smaller weight package. This release uses FP8 E4M3 weights in 128 × 128 blocks, with dynamic FP8 activations. It retains the full recurrent-state, local-attention, and Rock Memory Anchors (RMA) architecture.

The source is Rock Matrix's EdgeRock, frozen at a6a53a5573f1. Quantization changes the large projection and MLP matrices. RMA weights, small recurrent gates, the tied embedding/output head, cache capacity, and recurrent-state settings retain their source values. This is mixed-precision quantization; the format name describes the quantized matrices.

Load and generate

hf download rockmatrix-ai/EdgeRock-4B-FP8 --local-dir ./EdgeRock-4B-FP8
cd EdgeRock-4B-FP8
python -m pip install -r requirements.txt

Use a fresh Python environment and install this repository's requirements.txt. The pinned Transformers wheel contains the Qwen3.5 implementation used by EdgeRock. This package contains the text model and supports the PyTorch inference path.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "rockmatrix-ai/EdgeRock-4B-FP8"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
from transformers.integrations import finegrained_fp8
finegrained_fp8._deepgemm_available = False  # Validated Triton path for the pinned Transformers version.
model = AutoModelForCausalLM.from_pretrained(
    model_id, trust_remote_code=True, dtype=torch.bfloat16, device_map={"": 0},
).eval()
messages = [{"role": "user", "content": "Calculate 17 * 23."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt", add_special_tokens=False).to("cuda")
output = model.generate(**inputs, max_new_tokens=2048, do_sample=False)
print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))

FP8 execution requires an NVIDIA GPU with compute capability 9.0 or newer. The first run downloads the finegrained-FP8 kernel from Hugging Face.

The quantization configuration is embedded in config.json; loading does not require a second quantization pass.

Conversion validation

Check	BF16 source	This release
Weight files	8.527 GiB	5.208 GiB
Teacher-forced NLL on 32 GSM8K test examples	0.9892	0.9932
Maximum logit change after saving and reloading	—	0

The NLL check scores the final up-to-64 tokens of each question/answer sequence, truncated to 512 tokens. Lower is better; these examples are held out from conversion and are not calibration data. This is a conversion check, not a GSM8K accuracy score.

Fresh-process checks verified 168 protected tensors against the source bit for bit. All three arithmetic checks and both identifier-retrieval checks passed; retrieval prompts span approximately 2K and 8K tokens. A Chinese-to-English translation check is included. See validation.json for the exact protocol and outputs. Full benchmark accuracy and optimized serving throughput have not been remeasured for these quantized weights.

A separate generation screen uses GSM8K test rows 32–47, greedy decoding and a 4,096-token output budget. Correct complete answers: 15/16, compared with 15/16 for the BF16 source. Two additional memory update/key-disambiguation checks use 9,664 and 10,462 input tokens: 2/2, compared with 2/2 for BF16. A correct answer must match the expected value, close the thinking section, and reach EOS within the budget. These small screens do not establish unchanged full-benchmark accuracy.

The source model's multimodal frontend and native serving adapters are distributed in the original repository. This quantized package validates text inference; it does not bundle or validate the visual frontend. It is not a GGUF or Ollama model.

License and provenance

The weights retain the EdgeRock Model License 1.0; code retains Apache-2.0. Original notices are preserved in NOTICE. Changes in this derivative release consist of weight quantization, the packaged chat template and generation settings, and the packaged inference dependencies. Source benchmark claims are described in the technical report.