Library · Edge AI Radar · 2026-10-08

Daily

Which fresh models actually fit your board.

A daily, deterministic scan of new GGUF and ONNX releases on Hugging Face — verified file bytes in, datasheet memory ceilings out. Every fit verdict on this page is arithmetic you can check, not a benchmark we invented.

Memory landscape ↓Fit matrix ↓BPW spectrum ↓Methodology ← Library

Thursday, October 8, 2026 · 12 models × 10 boards · generated Oct 8, 07:23 PM UTC

Daily Snapshot · 2026-10-08

Real file sizes vs real hardware memory limits.

Quantized model weights vary dramatically by quantization scheme. Below is today's verified fit profile across microcontrollers, single-board computers, and edge GPU accelerators.

Scanned

HF Releases · 10 boards

Ceilings

7M–30G

RAM Span · MCU→Sigma

SBC+GPU Fits

12/12

Any ≥6.5 GiB board

16 GB+ Fits

12/12

Pi5 16GB / Orin NX · Sigma 12/12

Interactive Board Coverage Radar

Click any board vertex or pill below to inspect model headroom & RAM limits.

Fits (≤80%) Tight (≤100%)

sbc

LattePanda Sigma (32 GB)

Ceiling: 30.0 GiB32 GB LPDDR5 (x86)

✔ 12 Fits✖ 0 Exceed Ceiling

Mellum2.1-12B-A2.5B-Thinking-GGUF(10.3 GiB)d1-omni-600M-ONNX(537 MiB)d1-3B-ONNX(0 MiB)granite-3.0-2b-instruct(8 MiB)

Interactive Simulator 01

Memory topology & context pressure

Simulate how model tensor weights, runtime scratchpads, and expanding KV cache context windows occupy memory across hardware targets — from 7 MiB microcontrollers to 30 GiB x86 edge servers (Pi 5 8GB/16GB, Orange Pi 5 Plus 16GB, Rock 5B 16GB, Orin Nano/NX, Sigma 32GB).

Live Context Simulator

Hardware RAM Containment & Context Pressure

Pick a real Hugging Face release (verified file bytes). The bar showsweights + 12% overhead + KV cache growing with context. MCU & Coral TPU rows show architectural class mismatch for 2–16 GiB LLMs — not a fabricated 98,000% overflow.

Weights KV Cache (context) Overhead Tight Exceeds

1. Select model (verified bytes)12 releases today

2. Context window2,048 tokens

5122k4k8k

KV cache grows linearly with tokens. At — vs 512 baseline.Formula: max(runtime×ratio×seq/2048, 256 KiB). ratio 0.08 MCU, 0.04 SBC/GPU.

Active:bartowski/Mellum2.1-12B-A2.5B-Thinking-GGUF(gguf · Q4_K_M)

Tensors 8.21 GiBOverhead 1009 MiBKV 421 MiBTotal 9.20 GiB

Single-Board

· 6.5–30 GiB model budget · 5 boards

Sigma 32GB32 GB LPDDR5 (x86)

Ceiling 30.0 GiB

Total 9.61 GiB (32%)KV 421 MiB

✅ Fits · 20.4 GiB free

Pi 5 16GB16 GB LPDDR4X

Ceiling 14.5 GiB

Total 9.61 GiB (66%)KV 421 MiB

✅ Fits · 4.89 GiB free

Orange Pi 5+ 16GB16 GB LPDDR4X (RK3588)

Ceiling 14.0 GiB

Total 9.61 GiB (69%)KV 421 MiB

✅ Fits · 4.39 GiB free

ROCK 5B 16GB16 GB LPDDR4X (RK3588)

Ceiling 14.0 GiB

Total 9.61 GiB (69%)KV 421 MiB

✅ Fits · 4.39 GiB free

Pi 5 8GB8 GB LPDDR4X

Ceiling 6.50 GiB

Total 9.61 GiB (148%)KV 421 MiB

🛑 Exceeds by +3.11 GiB (148%)

Edge GPU

· Unified RAM · CUDA/TensorRT · 2 boards

Orin NX 16GB16 GB LPDDR5 (100 TOPS)

Ceiling 14.5 GiB

Total 9.61 GiB (66%)KV 421 MiB

✅ Fits · 4.89 GiB free

Orin Nano 8GB8 GB LPDDR5 (unified)

Ceiling 6.50 GiB

Total 9.61 GiB (148%)KV 421 MiB

🛑 Exceeds by +3.11 GiB (148%)

AI Accelerator

· 8 MB cache · int8 graph · 1 board

Coral TPU8 MB on-chip model cache

Ceiling 8.0 MiB

Not in class · 8.0 MiB ceiling vs 8.21 GiB model

Edge TPU 8 MB cache — graph spill to host

Microcontroller

· 7 MiB arena · TinyML only · 2 boards

Teensy 4.1 8MB1 MB on-chip + 8 MB PSRAM

Ceiling 7.0 MiB

Not in class · 7.0 MiB ceiling vs 8.21 GiB model

MCU ceiling — needs <16 MiB TinyML model

ESP32-S3 N8R8512 KB SRAM + 8 MB PSRAM

Ceiling 7.0 MiB

Not in class · 7.0 MiB ceiling vs 8.21 GiB model

MCU ceiling — needs <16 MiB TinyML model

7 MiB MCU → 30 GiB 16/32 GB SBCs · KV linear with context · 12% overhead constantVerified bytes × 1.25 + KV(seq)

🟢 Weights + 🟦 Overhead + 🔷 KV Cache7M–30G MCU → Sigma 32GB512 to 8,192 token context scaling

Data Table 02

Today's fit matrix

Representative quant file per repo (Q4_K_M preferred), runtime estimate = verified bytes × 1.25, against each board's disclosed model-memory ceiling. Fits ≤ 80% of ceiling · Tight ≤ 100% · otherwise No.

Scan 2026-10-08

Model · representative file	Runtime est.	Sigma 32GB30.00 GiB ceiling	Orin NX 16GB14.50 GiB ceiling	Pi 5 16GB14.50 GiB ceiling	Orange Pi 5+ 16GB14.00 GiB ceiling	ROCK 5B 16GB14.00 GiB ceiling	Orin Nano 8GB6.50 GiB ceiling	Pi 5 8GB6.50 GiB ceiling	Coral TPU8.0 MiB ceiling	Teensy 4.1 8MB7.0 MiB ceiling	ESP32-S3 N8R87.0 MiB ceiling	Score
bartowski/Mellum2.1-12B-A2.5B-Thinking-GGUFGGUF · Q4_K_M · 8.21 GiB file · text-generation	10.27 GiB	Fits	Fits	Fits	Fits	Fits	No	No	No	No	No	5/10
onnx-community/d1-omni-600M-ONNXONNX · 429.5 MiB file	536.9 MiB	Fits	Fits	Fits	Fits	Fits	Fits	Fits	No	No	No	7/10
onnx-community/d1-3B-ONNXONNX · 0.3 MiB file · image-text-to-text	0.4 MiB	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	10/10
onnx-community/granite-3.0-2b-instructONNX · 6.0 MiB file · text-generation	7.5 MiB	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Tight	No	No	8/10
onnx-community/gpt-oss-20b-ONNXONNX · 52.6 MiB file · text-generation	65.8 MiB	Fits	Fits	Fits	Fits	Fits	Fits	Fits	No	No	No	7/10
onnx-community/gemma-4-E4B-it-qat-mobile-ONNXONNX · 1.0 MiB file · any-to-any	1.2 MiB	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	10/10
onnx-community/gemma-4-E4B-it-ONNXONNX · 0.8 MiB file · any-to-any	1.0 MiB	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	10/10
onnx-community/gemma-4-E2B-it-qat-mobile-ONNXONNX · 0.8 MiB file · any-to-any	1.0 MiB	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	10/10
onnx-community/gemma-4-E2B-it-ONNXONNX · 0.6 MiB file · any-to-any	0.8 MiB	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	10/10
onnx-community/Qwen3.5-4B-ONNXONNX · 1.4 MiB file	1.7 MiB	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	10/10
onnx-community/Qwen3.5-4B-ONNX-OPTONNX · 0.9 MiB file	1.1 MiB	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	Fits	10/10
bartowski/K2-Horizon-7B-GGUFGGUF · Q4_K_M · 5.38 GiB file · text-generation	6.72 GiB	Fits	Fits	Fits	Fits	Fits	No	No	No	No	No	5/10

Analytics 03 & 04

Quantization & board fit analytics

Bits-per-weight distribution across today's models and breakdown of fit capacity per hardware board class.

Quantization Spectrum · Bits Per Weight (bpw)

Lower bpw fits smaller boards; 4.85 bpw (Q4_K_M) retains ~99% perplexity with 70% RAM savings.

12 models across 2 quants

Board Fit Breakdown · 12 Scanned Models

Percentage of today's models fitting within each board's datasheet memory limit.

Fits (≤80%) Tight (80-100%) No (>100%)

sbc30.0 GiB

Sigma 32GB

12 Fits0 No

edge-gpu14.5 GiB

Orin NX 16GB

12 Fits0 No

sbc14.5 GiB

Pi 5 16GB

12 Fits0 No

sbc14.0 GiB

Orange Pi 5+ 16GB

12 Fits0 No

sbc14.0 GiB

ROCK 5B 16GB

12 Fits0 No

edge-gpu6.5 GiB

Orin Nano 8GB

10 Fits2 No

sbc6.5 GiB

Pi 5 8GB

10 Fits2 No

accelerator8 MiB

Coral TPU

7 Fits1 Tight4 No

mcu7 MiB

Teensy 4.1 8MB

7 Fits5 No

mcu7 MiB

ESP32-S3 N8R8

7 Fits5 No

Per-Model Cards 05

The arithmetic, per model

File bytes come from the repo's verified Hugging Face file tree. Parameter count is back-derived from those bytes ÷ the quant's published bits-per-weight. FLOPs/token is the dense-transformer 2·P rule.

gguf · Q4_K_M5/10 boards
bartowski/Mellum2.1-12B-A2.5B-Thinking-GGUFMellum2.1-12B-A2.5B-Thinking-Q4_K_M.gguf
File
8.21 GiB
Runtime est.
10.27 GiB
≈ Params
14.55B
FLOPs/token
29.1 GFLOP
Footprint vs ceilings · log 1 MiB→32 GiB10.27 GiB
onnx7/10 boards
onnx-community/d1-omni-600M-ONNXonnx/audio_encoder.onnx
File
429.5 MiB
Runtime est.
536.9 MiB
≈ Params
—
FLOPs/token
—
Footprint vs ceilings · log 1 MiB→32 GiB536.9 MiB
onnx10/10 boards
onnx-community/d1-3B-ONNXonnx/decoder_model_merged_quantized.onnx
File
0.3 MiB
Runtime est.
0.4 MiB
≈ Params
—
FLOPs/token
—
Footprint vs ceilings · log 1 MiB→32 GiB0.4 MiB
onnx8/10 boards
onnx-community/granite-3.0-2b-instructonnx/model_quantized.onnx
File
6.0 MiB
Runtime est.
7.5 MiB
≈ Params
—
FLOPs/token
—
Footprint vs ceilings · log 1 MiB→32 GiB7.5 MiB
onnx7/10 boards
onnx-community/gpt-oss-20b-ONNXonnx/model_q4f16.onnx
File
52.6 MiB
Runtime est.
65.8 MiB
≈ Params
—
FLOPs/token
—
Footprint vs ceilings · log 1 MiB→32 GiB65.8 MiB
onnx10/10 boards
onnx-community/gemma-4-E4B-it-qat-mobile-ONNXonnx/decoder_model_merged_q2f16.onnx
File
1.0 MiB
Runtime est.
1.2 MiB
≈ Params
—
FLOPs/token
—
Footprint vs ceilings · log 1 MiB→32 GiB1.2 MiB
onnx10/10 boards
onnx-community/gemma-4-E4B-it-ONNXonnx/decoder_model_merged_q4f16.onnx
File
0.8 MiB
Runtime est.
1.0 MiB
≈ Params
—
FLOPs/token
—
Footprint vs ceilings · log 1 MiB→32 GiB1.0 MiB
onnx10/10 boards
onnx-community/gemma-4-E2B-it-qat-mobile-ONNXonnx/decoder_model_merged_q2f16.onnx
File
0.8 MiB
Runtime est.
1.0 MiB
≈ Params
—
FLOPs/token
—
Footprint vs ceilings · log 1 MiB→32 GiB1.0 MiB
onnx10/10 boards
onnx-community/gemma-4-E2B-it-ONNXonnx/decoder_model_merged_q4f16.onnx
File
0.6 MiB
Runtime est.
0.8 MiB
≈ Params
—
FLOPs/token
—
Footprint vs ceilings · log 1 MiB→32 GiB0.8 MiB
onnx10/10 boards
onnx-community/Qwen3.5-4B-ONNXonnx/decoder_model_merged_q4f16.onnx
File
1.4 MiB
Runtime est.
1.7 MiB
≈ Params
—
FLOPs/token
—
Footprint vs ceilings · log 1 MiB→32 GiB1.7 MiB
onnx10/10 boards
onnx-community/Qwen3.5-4B-ONNX-OPTonnx/decoder_model_merged_q4f16.onnx
File
0.9 MiB
Runtime est.
1.1 MiB
≈ Params
—
FLOPs/token
—
Footprint vs ceilings · log 1 MiB→32 GiB1.1 MiB
gguf · Q4_K_M5/10 boards
bartowski/K2-Horizon-7B-GGUFK2-Horizon-7B-Q4_K_M.gguf
File
5.38 GiB
Runtime est.
6.72 GiB
≈ Params
9.53B
FLOPs/token
19.1 GFLOP
Footprint vs ceilings · log 1 MiB→32 GiB6.72 GiB

Historical Trend 06

Daily release & fit history

Tracking total Hugging Face releases vs edge-fittable models across daily radar snapshots.

Edge AI Release & Fit Trend · 77 Daily Snapshots

Daily count of fresh Hugging Face model releases and how many fit edge hardware ceilings.

Total Scanned Edge Fittable

Hardware 07

The boards behind the ceilings

Prices shown were retrieved from the Amazon Product Advertising API on 8 October 2026 and are indicative only — the price and availability on Amazon at the time of purchase apply.

sbc30.00 GiB ceiling

Sigma 32GB

32 GB LPDDR5 (x86)

x86 edge server — 30 GiB model ceiling fits Gemma 4-26B-A4B Q4_K_M (15.8 GiB file / 19.7 GiB runtime) with 10 GiB headroom for 8k context.

LattePanda Sigma - 32GB (DFRobot DFR1080)$729 ↗

edge-gpu14.50 GiB ceiling

Orin NX 16GB

16 GB LPDDR5 (100 TOPS)

Orin NX 16GB — double RAM vs Nano, fits 4-12B Q4_K_M with long context (8k) headroom for TensorRT-LLM.

Yahboom Jetson Orin NX Super 16GB RAM 157 Tops Developer Kit Ubuntu Jetpack6.2 with 256GB SSD, Power Supply, for AI Large Model (Orin NX 16GB Developer Kit)↗

sbc14.50 GiB ceiling

Pi 5 16GB

16 GB LPDDR4X

16GB flagship — fits 9-12B Q4_K_M (5-7 GiB files) with 2-4k context headroom. Same power envelope, double RAM for local LLM.

Seeed Studio Raspberry Pi 5 Starter Kit - 16GB RAM, 64GBGB Micro SD Card Pre-Loaded 64-bit OS, Type-C Power Supply, Active Cooling Case for Edge AI, Docker & Pro Workstation↗

sbc14.00 GiB ceiling

Orange Pi 5+ 16GB

16 GB LPDDR4X (RK3588)

RK3588 SBC — cheaper Pi5-class alternative with PCIe NVMe for model storage; 14 GiB ceiling after RKLLM / llama.cpp overhead.

Orange Pi 5 Plus 16GB Rockchip RK3588 8 Core 64 Bit Single Board Computer, 2.4GHz Frequency 8K Video Decoding Open Source Development Board Run Orange Pi OS, Android, Debian, Ubuntu (5 Plus 16G)$353.99 ↗

sbc14.00 GiB ceiling

ROCK 5B 16GB

16 GB LPDDR4X (RK3588)

ROCK 5B 16GB — solid mid-range for 7-12B Q4_K_M; active cooling recommended for sustained inference.

Radxa ROCK 5B - 16GB$159 ↗

edge-gpu6.50 GiB ceiling

Orin Nano 8GB

8 GB LPDDR5 (unified)

Unified memory — CUDA/TensorRT runtime, desktop and display all share the same 8 GB pool.

sbc6.50 GiB ceiling

Pi 5 8GB

8 GB LPDDR4X

Full Linux host — llama.cpp / ONNX Runtime with headroom for the OS and KV cache growth at long context.

CanaKit Raspberry Pi 5 Starter Kit PRO - Turbine Black (128GB Edition) (8GB RAM)$259.95 ↗

accelerator8.0 MiB ceiling

Coral TPU

8 MB on-chip model cache

Full-speed int8 only while the compiled model stays inside the 8 MB cache; larger graphs spill to host RAM. Host supplies its own memory via USB.

Google Coral USB Edge TPU ML Accelerator coprocessor for Raspberry Pi and Other Embedded Single Board Computers↗

mcu7.0 MiB ceiling

Teensy 4.1 8MB

1 MB on-chip + 8 MB PSRAM

Ceiling assumes the PSRAM fitted. On-chip alone caps models near 0.75 MB. Linked card is the Teensy 4.0 sibling (same 600 MHz i.MX RT1062 core; the 4.1 adds the PSRAM pads this ceiling needs).

Teensy 4.0$23.80 ↗

mcu7.0 MiB ceiling

ESP32-S3 N8R8

512 KB SRAM + 8 MB PSRAM

Model arena lives in octal PSRAM; on-chip SRAM stays free for the RTOS and tensor arena scratch.

ESP32-S3-DevKitC-1-N8R8 Development Board↗

Verification 08

Methodology — deterministic, checkable

Inputs (ground truth only)

Verified file bytes — each repo's public Hugging Face file tree, fetched at build time. No benchmark claims, no tokens/sec, no accuracy numbers.
Datasheet memory constants — ESP32-S3 N8R8 (512 KB + 8 MB PSRAM), Teensy 4.1 +8MB, Pi 5 8GB/16GB, Orange Pi 5+ 16GB, ROCK 5B 16GB, Coral TPU (8 MB cache), Orin Nano 8GB / NX 16GB, Sigma 32GB. Hardcoded in scripts/edgespec/radar-core.mjs with sources in comments.
Published quant formats — llama.cpp bits-per-weight values (Q4_K_M ≈ 4.85 bpw, Q8_0 ≈ 8.5, …).

Formulas (nothing else)

runtime = fileBytes × 1.25 (weights + KV/buffers)
≈params = fileBytes × 8 ÷ bitsPerWeight (GGUF only)
FLOPs/token = 2 × params (dense transformer rule)
fits ≤ 80% ceiling · tight ≤ 100% · else no

Gate tests (scripts/edgespec/pipeline.test.mjs) run before every publish in CI. Days with zero grounded models publish nothing — the same anti-abuse posture as the Signals Journal.

Questions 09

Frequently Asked Questions

Why no tokens-per-second or accuracy numbers?+

Because we'd have to invent them. Throughput depends on your exact runtime, cooling, and build flags — a number we didn't measure on the named board would be fabrication. This radar answers only the question arithmetic can answer honestly: does the model's memory footprint fit the board's ceiling?

Why does a “fits” verdict stop at 80% of the ceiling?+

KV cache grows with context length, runtimes allocate scratch buffers, and OS/framework overhead varies. The 20% margin is the difference between “loads once in a demo” and “runs reliably on your bench.” Tight means it fits on paper — plan your context budget carefully.

The repo name says 27B but your table shows fewer params — why?+

We derive ≈params from the verified file bytes ÷ the quant's bits-per-weight, not from the repo's title. Some repos ship partial, draft, or experimental files whose real size doesn't match the name. The bytes are the ground truth; the name is marketing.

When does this page update?+

A GitHub Action (edgespec-digest.yml) rescans Hugging Face, re-runs the gate tests, and commits a new dated snapshot only when grounded models pass. Its daily schedule is paused during an October 2026 site review, so the snapshot date shown is the last run. This site itself makes zero runtime requests — the fetch happens at build time, per our privacy boundary.

Which fresh models actually fit your board.

Real file sizes vs real hardware memory limits.

LattePanda Sigma (32 GB)

Jetson Orin NX (16 GB)

Raspberry Pi 5 (16 GB)

Orange Pi 5 Plus (16 GB)

Radxa ROCK 5B (16 GB)

Jetson Orin Nano (8 GB)

Raspberry Pi 5 (8 GB)

Coral Edge TPU (USB Accelerator)

Teensy 4.1 (+8 MB PSRAM)

ESP32-S3 (N8R8 devkit)

Memory topology & context pressure

Single-Board

Edge GPU

AI Accelerator

Microcontroller

Today's fit matrix

Quantization & board fit analytics

Quantization Spectrum · Bits Per Weight (bpw)

Board Fit Breakdown · 12 Scanned Models

Sigma 32GB

Orin NX 16GB

Pi 5 16GB

Orange Pi 5+ 16GB

ROCK 5B 16GB

Orin Nano 8GB

Pi 5 8GB

Coral TPU

Teensy 4.1 8MB

ESP32-S3 N8R8

The arithmetic, per model

Daily release & fit history

Edge AI Release & Fit Trend · 77 Daily Snapshots

The boards behind the ceilings

Sigma 32GB

Orin NX 16GB

Pi 5 16GB

Orange Pi 5+ 16GB

ROCK 5B 16GB

Orin Nano 8GB

Pi 5 8GB

Coral TPU

Teensy 4.1 8MB

ESP32-S3 N8R8

Methodology — deterministic, checkable

Inputs (ground truth only)

Formulas (nothing else)

Frequently Asked Questions