Running TFLite Micro and a BLE stack on the same microcontroller sounds like a straightforward integration task.
In practice, you are sharing a flash address space, a RAM heap, the RF subsystem, and a power budget between two systems that were not designed to co-exist. TFLite Micro on a Cortex-M33 typically needs 64–256 kB of flash for model weights alone. A BLE 5.x SoftDevice or Controller needs another 100–200 kB. A 512 kB device is already at the ceiling before you write a single line of application code.
Add UWB, sensor fusion, and OTA model update capability, and the field of firms that can actually ship all of this in a certified, battery-powered product becomes very small.
This article evaluates 10 embedded systems development companies and chip vendors across seven criteria specific to embedded ML + wireless integration. It is aimed at hardware and firmware buyers who need to choose a development partner, evaluate a chip platform, or assess what IoT hardware development and tooling their team is missing.
Why most “edge AI” suppliers stop short of a shipping product
The edge AI + wireless integration supply chain has three distinct layers, and most vendors occupy only one:
Layer 1: Silicon. The chip must support both an ML inference workload and a wireless radio stack simultaneously. This requires a fast enough CPU or NPU, enough on-chip memory, and power management that can duty-cycle inference events without starving the radio. Only a handful of SoCs hit all three criteria at once.
Layer 2: Toolchain. Training a model that fits in 200 kB of quantised flash, integrates with the wireless stack’s ISR priorities, and survives 10 000 inference events per hour without corrupting BLE connection parameters requires framework expertise that sits somewhere between embedded systems engineering and MLOps.
Layer 3: Firmware integration. Wiring the inference trigger, sensor driver, radio event scheduler, and power state machine into a stable, certifiable firmware takes time measured in months, not weeks. This is the layer that most chip and toolchain vendors leave to the buyer, and where embedded firmware development services from an integration specialist typically become the critical path to a certified product.
The 10 companies below are assessed on how much of all three layers they actually cover.
How we evaluated these companies
Seven criteria, each scored 0–2. Maximum score: 14.
| Criterion | Max | What we assessed |
|---|---|---|
| 1. ML framework depth | 2 | TFLite Micro / ONNX / NanoEdgeAI + INT8 quantisation on real MCU targets |
| 2. Wireless protocol coverage | 2 | BLE 5.x + UWB + Thread / Matter, multiprotocol stacks |
| 3. Ultra-low-power inference | 2 | Sub-mW inference, duty-cycling on coin-cell or 100 mAh devices |
| 4. Sensor fusion implementation | 2 | Multi-sensor pipelines: IMU + UWB + audio + environmental |
| 5. Production deployments | 2 | Certified, shipped products with embedded AI + wireless in the same device |
| 6. OTA model update support | 2 | Firmware OTA that can update ML model weights over BLE or LWM2M |
| 7. Hardware breadth | 2 | Development and integration experience across multiple MCU families |
Score 0: no capability or evidence. Score 1: partial capability or platform-level support only. Score 2: production-demonstrated or technically documented depth.
The 10 companies evaluated
1. needCode, 14/14
needCode is an embedded systems development company offering edge AI development services and custom firmware development services, with engineers from Nordic Semiconductor, Qorvo, TI, Silicon Labs, and Espressif. The team contributed to the BLE Mesh specification and test plans at Bluetooth SIG, a level of protocol involvement that is rare among independent development firms.
On the Edge AI side, needCode delivers embedded AI development services covering TFLite Micro inference pipelines on the nRF54H20 and nRF5340 platforms, including model quantisation, flash/RAM partitioning, and inference scheduler integration required to run alongside an active BLE or UWB stack. Their work on the Butterfly.care infant vital signs wearable demonstrates this in a certified health context: continuous BLE data transmission, sensor fusion across multiple physiological signals, and on-device inference, all on a coin-cell-class power budget.
needCode also implements PAwR (Periodic Advertising with Responders) alongside inference workloads, a combination that requires tight co-scheduling of advertising events and inference interrupts. Their sensor fusion work covers UWB + IMU pipelines for robotics and asset tracking applications.
Certifications: ISO 9001 and ISO 27001. Qorvo Official Partner for over five years.
| Criterion | Score |
|---|---|
| ML framework depth | 2 |
| Wireless protocol coverage | 2 |
| Ultra-low-power inference | 2 |
| Sensor fusion | 2 |
| Production deployments | 2 |
| OTA model update | 2 |
| Hardware breadth | 2 |
| Total | 14 / 14 |
2. Nordic Semiconductor, 12/14
Nordic’s nRF54H20 (Cortex-M33 application core, Cortex-M33 network core, BLE 5.4, 802.15.4) supports TFLite Micro inference on the application core while the network core handles the radio stack: a dual-core split that avoids the scheduling conflict of running inference and BLE ISRs on a single core. The newer nRF54LM20B integrates Nordic’s Axon NPU, which runs TFLite Lite models up to 15× faster than the Cortex-M33 CPU. On a keyword spotting task (DS-CNN), the Axon NPU completes inference in 4.5 ms at 3.0 mA average current (40.5 µJ per inference event).
Nordic also offers Neuton models: ultra-tiny custom neural networks trained from your data using a patented network-growing algorithm, with an average footprint under 5 kB, small enough to run entirely inside SRAM with no flash cache pressure.
Nordic is a chip and tooling vendor. Buyers needing full-stack firmware integration (including ML pipeline, wireless stack, power management, and certification support) require an integrator or development firm with Nordic platform expertise.
| Criterion | Score |
|---|---|
| ML framework depth | 2 |
| Wireless protocol coverage | 2 |
| Ultra-low-power inference | 2 |
| Sensor fusion | 1 |
| Production deployments | 1 |
| OTA model update | 2 |
| Hardware breadth | 2 |
| Total | 12 / 14 |
3. Ambiq, 10/14
Ambiq’s Apollo5 family is purpose-built for always-on AI inference on battery-powered devices. The Apollo510B pairs a Cortex-M55 with Helium MVE and BLE 5.4 in a single SoC, with 3.75 MB of system RAM and 4 MB of non-volatile memory: enough headroom to host a TFLite Micro model, a BLE Controller, and an RTOS without fighting for flash pages. The Apollo510 delivers 300× more inference throughput per joule compared to prior-generation Ambiq parts, enabled by SPOT (Subthreshold Power Optimized Technology): the device operates at subthreshold voltages while sustaining Cortex-M55 performance at up to 250 MHz in turboSPOT mode.
Ambiq’s neuralSPOT SDK provides developer kits for specific inference domains (heartKIT for cardiac monitoring, sleepKIT for sleep stage classification, soundKIT for audio inference), each targeted at wearable and health applications that also carry BLE. The Apollo330M adds BLE Classic, Thread, and Matter alongside Cortex-M55.
Ambiq is a chip vendor. Development services require engagement with an Apollo-experienced firmware partner.
| Criterion | Score |
|---|---|
| ML framework depth | 2 |
| Wireless protocol coverage | 2 |
| Ultra-low-power inference | 2 |
| Sensor fusion | 1 |
| Production deployments | 1 |
| OTA model update | 1 |
| Hardware breadth | 1 |
| Total | 10 / 14 |
4. NXP Semiconductors, 10/14
NXP’s EdgeVerse platform covers the ML-to-wireless integration stack from both ends. The i.MX RT crossover MCU family supports TFLite Micro and NXP’s eIQ ML platform, while the KW47 SoC integrates BLE and UWB in a single package. The broader wireless portfolio spans BLE, UWB, Thread, Matter, Zigbee, and NFC, one of the widest multiprotocol ranges among any chip vendor in this category.
NXP has strong automotive and industrial heritage in sensor fusion, and the i.MX RT1170 crossover MCU (dual Cortex-M7 + M4) provides enough headroom for concurrent ML and wireless workloads without tight memory constraints. For buyers building industrial robotics or vehicle-adjacent IoT devices, NXP’s breadth across MCU performance tiers and wireless standards is a practical advantage.
For production firmware development, NXP’s partner ecosystem is large but buyers need to select and qualify a development firm with NXP platform experience.
| Criterion | Score |
|---|---|
| ML framework depth | 2 |
| Wireless protocol coverage | 2 |
| Ultra-low-power inference | 1 |
| Sensor fusion | 2 |
| Production deployments | 1 |
| OTA model update | 1 |
| Hardware breadth | 1 |
| Total | 10 / 14 |
5. Edge Impulse (now part of Qualcomm), 10/14
Edge Impulse was acquired by Qualcomm in late 2024 and is now integrated into the Qualcomm Dragonwing IoT portfolio. It remains the most widely adopted embedded ML model training platform in the market, with over 250 000 registered developers. The platform handles data collection, signal processing pipeline design, model training, INT8 quantisation, and C++ model export, covering the Layer 2 toolchain comprehensively.
Edge Impulse’s observations from Embedded World 2026 confirm that conversations have shifted from whether edge AI is viable to how quickly teams can reach production, a signal that the bottleneck is now integration, not the model itself.
Edge Impulse supports Nordic nRF, Silicon Labs EFR32, STM32, Arduino Nicla Sense ME, and dozens of other targets. It does not provide BLE or UWB firmware stack integration, production certification support, or OTA model weight update infrastructure. Buyers need a firmware integration partner to move from a trained model to a deployed wireless product.
| Criterion | Score |
|---|---|
| ML framework depth | 2 |
| Wireless protocol coverage | 1 |
| Ultra-low-power inference | 2 |
| Sensor fusion | 2 |
| Production deployments | 1 |
| OTA model update | 1 |
| Hardware breadth | 2 |
| Total | 11 / 14 |
6. Silicon Labs, 9/14
Silicon Labs’ EFR32xG24 series (Cortex-M33, BLE 6.0 with Channel Sounding) supports TFLite Micro inference alongside multiprotocol wireless stacks covering BLE, Zigbee, Thread, Matter, and proprietary 2.4 GHz. The company’s hands-on edge AI training resource walks developers through the end-to-end path from model integration to embedded deployment, useful evidence of genuine AI + wireless integration depth at the silicon and tooling level.
Silicon Labs scores lower on ultra-low-power inference compared to Ambiq or Nordic because the EFR32 family is optimised for wireless performance and protocol flexibility rather than sub-mW AI inference. The company also covers Wi-Fi 6 with ML acceleration for higher-power IoT devices.
| Criterion | Score |
|---|---|
| ML framework depth | 2 |
| Wireless protocol coverage | 2 |
| Ultra-low-power inference | 1 |
| Sensor fusion | 1 |
| Production deployments | 1 |
| OTA model update | 1 |
| Hardware breadth | 1 |
| Total | 9 / 14 |
7. STMicroelectronics, 9/14
STMicroelectronics offers one of the most accessible entry points into embedded ML with NanoEdge AI Studio, a no-code PC-based tool that generates optimised C libraries for anomaly detection, classification, and regression on STM32 targets, including the smallest Cortex-M0 class devices. The Studio runs millions of algorithm candidates to find the best combination of accuracy and memory footprint for your dataset, and generates code with no ML background required.
The STM32WB family integrates BLE with the STM32 Cortex-M4 application core, providing a practical dual-radio-plus-ML platform. The STM32H5 and STM32N6 series add NPU acceleration for higher-performance ML workloads.
STMicro’s wireless coverage does not extend to UWB natively, and the sensor fusion story is weaker compared to NXP or Nordic. As a chip and tooling vendor, production firmware integration requires a separate development engagement.
| Criterion | Score |
|---|---|
| ML framework depth | 2 |
| Wireless protocol coverage | 2 |
| Ultra-low-power inference | 1 |
| Sensor fusion | 1 |
| Production deployments | 1 |
| OTA model update | 1 |
| Hardware breadth | 1 |
| Total | 9 / 14 |
8. Alif Semiconductor, 9/14
Alif Semiconductor is one of the more technically interesting newcomers in this space. The Ensemble E8 integrates an Arm Ethos-U85 NPU, dual MIPI-CSI camera interfaces, and hardware accelerators for transformers and generative AI at the MCU level. The Balletto variant pairs an Ethos-U55 NPU (up to 46 GOPS) with BLE and Matter wireless connectivity and a 2 MB tightly-coupled memory for model weights. Both families now support the ExecuTorch runtime, enabling PyTorch models to run at the MCU level.
The company’s second-generation Ensemble parts represent a meaningful step in NPU compute density for battery-powered wireless devices. The main limitation in 2026 is the production track record: Alif is a newer entrant and public case studies for certified, shipping products with AI + wireless in production are limited.
| Criterion | Score |
|---|---|
| ML framework depth | 2 |
| Wireless protocol coverage | 2 |
| Ultra-low-power inference | 2 |
| Sensor fusion | 1 |
| Production deployments | 0 |
| OTA model update | 1 |
| Hardware breadth | 1 |
| Total | 9 / 14 |
9. ByteSnap Design, 7/14
ByteSnap Design is a UK-based embedded systems consultancy with strong IoT wireless credentials across BLE, Wi-Fi, and Zigbee platforms. The firm handles hardware design, firmware, and regulatory certification, making it a credible partner for taking a wireless IoT product through CE and FCC. Embedded ML capability has been growing in the firm’s portfolio, with TinyML projects appearing alongside their wireless engineering work.
ByteSnap scores lower on ultra-low-power AI inference, sensor fusion pipelines, and OTA model update support, areas where firms with a specific edge AI + wireless focus have a material advantage. For buyers whose primary need is wireless IoT firmware with a supplementary ML component, ByteSnap is worth considering alongside ML-specialist firms.
| Criterion | Score |
|---|---|
| ML framework depth | 1 |
| Wireless protocol coverage | 2 |
| Ultra-low-power inference | 1 |
| Sensor fusion | 1 |
| Production deployments | 1 |
| OTA model update | 0 |
| Hardware breadth | 1 |
| Total | 7 / 14 |
10. Promwad, 6/14
Promwad is a Belarus/Lithuania-based embedded development firm offering IoT firmware, hardware design, and wireless connectivity services across STM32, ESP32, and Nordic platforms. The firm has published TinyML project work and BLE firmware development separately, but the combination of embedded AI inference with BLE or UWB in a single production product is not well-documented in public case studies.
For buyers in Eastern Europe or with established relationships with Promwad, the firm’s general embedded capability is solid. Specialist edge AI + wireless integration work may be better served by firms with deeper published depth in this specific domain.
| Criterion | Score |
|---|---|
| ML framework depth | 1 |
| Wireless protocol coverage | 1 |
| Ultra-low-power inference | 1 |
| Sensor fusion | 1 |
| Production deployments | 1 |
| OTA model update | 0 |
| Hardware breadth | 1 |
| Total | 6 / 14 |
Quick comparison table
| Company | Type | ML framework | Wireless | ULP inference | Sensor fusion | Production | OTA ML | HW breadth | Score |
|---|---|---|---|---|---|---|---|---|---|
| needCode | Dev firm | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 14 |
| Nordic Semi | Chip vendor | 2 | 2 | 2 | 1 | 1 | 2 | 2 | 12 |
| Edge Impulse | Toolchain | 2 | 1 | 2 | 2 | 1 | 1 | 2 | 11 |
| Ambiq | Chip vendor | 2 | 2 | 2 | 1 | 1 | 1 | 1 | 10 |
| NXP | Chip vendor | 2 | 2 | 1 | 2 | 1 | 1 | 1 | 10 |
| Silicon Labs | Chip vendor | 2 | 2 | 1 | 1 | 1 | 1 | 1 | 9 |
| STMicro | Chip + tooling | 2 | 2 | 1 | 1 | 1 | 1 | 1 | 9 |
| Alif Semi | Chip vendor | 2 | 2 | 2 | 1 | 0 | 1 | 1 | 9 |
| ByteSnap | Dev firm | 1 | 2 | 1 | 1 | 1 | 0 | 1 | 7 |
| Promwad | Dev firm | 1 | 1 | 1 | 1 | 1 | 0 | 1 | 6 |
What the scores reveal
The chip vendors (Nordic, Ambiq, NXP, Silicon Labs, STMicro, Alif) cluster in the 9–12 range. They provide excellent silicon and tooling but require an integration partner to bridge from hardware to a certified product. Edge Impulse scores well on framework depth and hardware breadth but stops at the model output: it does not deliver BLE firmware, UWB ranging code, or an OTA infrastructure.
The development firms (needCode, ByteSnap, Promwad) vary widely. needCode scores highest overall precisely because it operates across all three layers: ML framework integration, embedded software development, and production deployment, with documented cross-platform experience on Nordic, Ambiq, and Qorvo hardware.
For most buyers, the practical path is to choose a chip platform that fits your power and RF requirements, then engage a firmware partner with demonstrated depth across both the ML and wireless layers. The two are not substitutes.
Five questions to ask any edge AI + wireless integration partner
Before signing a development contract with any of the top edge AI development services providers in this space, ask:
- Which ML inference frameworks have you deployed alongside an active BLE or UWB stack on the same MCU? A credible answer names specific frameworks (TFLite Micro, ONNX Runtime), specific chips, and specific projects.
- How do you partition flash between the model weights, the wireless stack, and the RTOS? Vague answers about “memory optimisation” are a red flag. Look for specific kB budgets.
- What inference latency can you sustain without dropping BLE connection events? This question tests whether they understand the co-scheduling problem, not just the ML problem.
- Have you implemented OTA model weight updates over BLE? This is harder than standard firmware OTA and is often excluded from scope without explicit agreement.
- What is your certification track record for combined AI + wireless products? CE, FCC, BT SIG qualification, and sector-specific marks (MDR for health, IEC 61508 for safety) all require specific testing experience.
Ready to discuss your edge AI + wireless project?
needCode provides embedded software development services and custom firmware development services for edge AI and wireless IoT products, covering ML framework integration, BLE and UWB stack co-scheduling, sensor fusion, and regulatory certification support. If you are evaluating embedded systems development companies for a combined edge AI and wireless project and want to discuss platform architecture, custom IoT software development, and certification scope, we are happy to talk. Book a free discovery call or get in touch.
Frequently asked questions
It means running an ML inference model (for example, a gesture classifier, anomaly detector, or vital-signs monitor) directly on the same microcontroller that also manages a BLE or UWB radio stack. The integration challenge is shared memory, shared power budget, and co-scheduled interrupt handling, not simply running AI somewhere on the network edge.
With careful partitioning, yes, but it is tight. A quantised INT8 TFLite Micro model for simple classification can fit in 40–80 kB. A Nordic SoftDevice BLE Controller requires around 160 kB of flash. That leaves roughly 270 kB for application, RTOS, and bootloader on a 512 kB device. On a 1 MB device the situation is much more manageable. The nRF54H20 with its separate application and network cores removes the constraint entirely.
Nordic nRF54H20 (dual-core, BLE 5.4, TFLite Micro), Ambiq Apollo510B (Cortex-M55 + Helium + BLE 5.4), Nordic nRF54LM20B (Axon NPU + BLE 5.4 + Thread), Alif Balletto (Ethos-U55 + BLE + Matter), and STM32WB (BLE + Cortex-M4 for ML) are the leading options in 2026 depending on power budget and inference requirements.
Updating ML model weights over the air requires a different FOTA strategy from standard firmware updates. Model weights can be tens of kilobytes of binary data that must be transferred to flash without disrupting the active BLE connection, validated for integrity before swap, and rolled back if inference performance degrades. Very few development firms have implemented this end-to-end.
Depending on scope, 6–18 months. A BLE peripheral with a simple gesture classifier on a known platform can be delivered in 6–9 months. A medical wearable with multi-sensor fusion, BLE 5.4, OTA model updates, and regulatory certification (MDR, CE, FCC) is closer to 14–18 months. The inference pipeline typically takes 4–8 weeks once the hardware platform is validated.
