Use Bolt 74B-A9B 24GB for a local laptop or workstation with at least 24 GB of usable GPU memory. NVIDIA positions this model for RTX Spark laptops, GeForce RTX PCs, DGX Spark, DGX Station and RTX PRO workstations, with local support in llama.cpp, Ollama, LM Studio, Unsloth and vLLM. Use the smaller Bolt E4B 4GB for embedded conversational or physical-agent projects, including Jetson, once its scheduled October 23, 2026 release is available.
Direct iPhone and Android deployment is not documented for either variant. The published material does not provide a mobile package for Bolt, nor support for Core ML, NNAPI, Apple or Qualcomm mobile NPUs. Treat a phone as an unverified target rather than promising that the model will run locally.
As an Amazon Associate I earn from qualifying purchases.
Which Nemotron 3.5 Bolt model should you deploy?
Bolt is not one universal edge build. The two announced variants target very different memory and hardware classes. Choosing the wrong one is the most common deployment mistake: a 74-billion-parameter model cannot be treated like a small embedded model simply because only 9 billion parameters are active for each token.
| Variant | Intended hardware | Memory class | Model and performance information | Documented runtimes | Availability |
|---|---|---|---|---|---|
| Bolt 74B-A9B 24GB | RTX Spark laptops, GeForce RTX PCs, DGX Spark, DGX Station and RTX PRO workstations | Designed to fit a 24 GB GPU-memory budget using native roughly 2-bit quantization | 74 billion total parameters, 9 billion active parameters; NVIDIA’s design target is at least 80 tokens per second | llama.cpp, Ollama, LM Studio, Unsloth and vLLM | Published for local deployment; verify the exact checkpoint and runtime version before shipping |
| Bolt E4B 4GB | GeForce RTX, Jetson and RTX/DGX Spark systems | 4 GB class | Smaller model for fast conversational and physical agents, such as local voice assistants, interactive characters and embedded interfaces | Final runtime list and release package should be checked when the model is published | Scheduled for October 23, 2026; it was not yet available on October 16, 2026 |
The 74B-A9B model is compressed from Nemotron 3.5 Super VL. Its “A9B” name indicates the active-parameter count, not a 9-billion-parameter memory requirement. The checkpoint still has to accommodate the full model representation, which is why the 24 GB target and the quantization format matter.
#1 Best Overall
- READY FOR ANYTHING – Dive headfirst into gaming on Windows 11 powered by the Intel Core i5 Processor 13450HX and an NVIDIA GeForce RTX 5050 Laptop GPU with a Max TGP of 115W and NVIDIA Advanced Optimus.
- SUBTLE STYLING – The TUF Gaming F16 maintains its classic design, boasting a subtle embossed TUF logo on its sleek cover.
- IMMERSIVE VISUALS – The TUF Gaming F16’s FHD+ 165Hz display with 100% sRGB color draws you into the action. Adaptive-Sync technology reduces lag, minimizes stuttering, and eliminates visual tearing for ultra-smooth gameplay.
- MILITARY GRADE DURABILITY – As a TUF gaming machine, the F16 has been rigorously tested to meet Military Grade testing standards, MIL-STD-810H. Rest easy knowing this laptop will operate at peak performance in harsh conditions.
- EFFICIENT COOLING – Equipped with 2nd Gen Arc Flow Fans, full-width heatsink, and full-width vent, the TUF Gaming F16 optimizes cooling performance without extra noise.
Hardware requirements for Bolt 74B-A9B
Start with usable GPU memory, not the product name
Before downloading anything, check the memory exposed to the runtime. NVIDIA’s target is at least 24 GB of usable GPU memory. “24 GB” on a specification sheet is not always 24 GB available to a model: the operating system, display server, drivers and other processes can reserve part of it. Leave practical headroom for the runtime, key-value cache, context and temporary buffers instead of planning to consume every last megabyte.
- Qualifying hosts named by NVIDIA include RTX Spark laptops, GeForce RTX PCs, DGX Spark, DGX Station and RTX PRO workstations.
- Many ordinary RTX laptops have less VRAM than this model’s stated footprint and therefore do not meet the published target.
- System RAM can help with offloading in some runtimes, but it does not turn an undersized GPU into a 24 GB device. Expect lower speed and more complex tuning if layers spill out of GPU memory.
- Power and cooling are part of the specification in practice. A laptop that reaches the memory target may still throttle under a sustained generation workload.
Use the published quantized checkpoint for local inference
NVIDIA publishes both quantized and BF16 checkpoints. The technical report identifies the IQ2_XXS quantized model as the local 24 GB target. BF16 is intended for customization and requires substantially more memory than the edge-oriented quantized build. Do not substitute the BF16 checkpoint on a 24 GB laptop and assume that the quantized footprint will remain unchanged.
What the 24 GB figure does and does not guarantee
Fitting the model is only the first gate. Your usable context length, batch size, concurrent requests, image or multimodal inputs, and runtime settings all affect memory. A configuration that loads successfully with a short prompt can fail when the context grows or a second request arrives. Record free memory after the model loads and again during the longest prompt your application will send.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Deploy Bolt 74B-A9B on a laptop or workstation
1. Confirm the target machine
- Identify the exact GPU model and the memory available to applications, rather than relying on the family name alone.
- Close unrelated GPU workloads and check that the machine can sustain the required power and cooling for a long generation session.
- Confirm the operating system and CPU architecture are supported by your chosen runtime. NVIDIA’s cited performance result used a Windows 11 ARM x64 RTX Spark laptop configuration, but that result is not a universal requirement or guarantee.
2. Select and review the checkpoint
- Obtain the NVIDIA-published Bolt 74B-A9B 24GB quantized checkpoint.
- Confirm that the file is the IQ2_XXS-oriented local build, not the BF16 customization checkpoint.
- Read the license terms before distributing the model or embedding it in a product. Keep the checkpoint name, quantization format and revision in your deployment record so that a later update can be reproduced.
3. Choose a runtime for the job
| Runtime | Best fit in the published guidance | Deployment consideration |
|---|---|---|
| llama.cpp | RTX Spark local inference | NVIDIA identifies this as the optimized path for the quantized local model. It is the natural choice when you want a self-contained desktop or laptop process. |
| Ollama | Convenient local application integration | Use the exact Bolt model package and format supported by the current release; do not assume every community package has the same quantization or tokenizer settings. |
| LM Studio | Graphical local testing on a PC | Useful for validating prompts and memory behavior interactively. Treat the selected model file and backend as part of the test configuration. |
| Unsloth | Workflows that include model customization | Use when your goal includes adaptation or training-related work rather than only serving the published quantized model. |
| vLLM | DGX Spark or a supported NVIDIA server/workstation deployment | Prefer this when you need a service endpoint, batching or multi-request serving and the hardware and model format are supported by the current vLLM release. |
Install the runtime using its current official documentation, then select the Bolt checkpoint through that runtime’s model-loading interface. No single universal install command is published, because the correct command varies by operating system, runtime release and checkpoint packaging.
Rank #2
- Beyond Performance: The Intel Core i5-13420H processor goes beyond performance to let your PC do even more at once. With a first-of-its-kind design, you get the performance you need to play, record and stream games with high FPS and effortlessly switch to heavy multitasking workloads like video, music and photo editing.
- AI-Powered Graphics: The state-of-the-art GeForce RTX 4050 graphics (194 AI TOPS) provide stunning visuals and exceptional performance. DLSS 3.5 enhances ray tracing quality using AI, elevating your gaming experience with increased beauty, immersion, and realism.
- Visual Excellence: See your digital conquests unfold in vibrant Full HD on a 15.6" screen, perfectly timed at a quick 165Hz refresh rate and a wide 16:9 aspect ratio providing 82.64% screen-to-body ratio. Now you can land those reflexive shots with pinpoint accuracy and minimal ghosting. It's like having a portal to the gaming universe right on your lap.
- Internal Specifications: 8GB DDR5 Memory (2 DDR5 Slots Total, Maximum 32GB); 512GB PCIe Gen 4 SSD
- Stay Connected: Your gaming sanctuary is wherever you are. On the couch? Settle in with fast and stable Wi-Fi 6. Gaming cafe? Get an edge online with Killer Ethernet E2600 Gigabit Ethernet. No matter your location, Nitro V 15 ensures you're always in the driver's seat. With the powerful Thunderbolt 4 port, you have the trifecta of power charging and data transfer with bidirectional movement and video display in one interface.
4. Enable speculative decoding when supported
NVIDIA supplies a LiLiCorr speculative-decoding drafter for Bolt 74B-A9B 24GB. The drafter has 1.13 billion parameters and is intended to reduce latency at low concurrency. It proposes tokens with a smaller model so the main model can verify several tokens efficiently. Whether it helps depends on the runtime, prompt mix, batch size and hardware; enable it only through a documented integration and measure the result against a baseline.
5. Establish a baseline before integrating an application
- Run a fixed prompt set with the same sampling settings you expect in production.
- Measure time to first token, steady-state tokens per second, total response time, peak GPU memory, context length and error rate.
- Repeat at the concurrency your application will actually use. The LiLiCorr drafter is specifically described as useful at low concurrency, so a high-concurrency service may show a different trade-off.
- Repeat the test after a sustained run to expose thermal throttling and power-limit behavior.
NVIDIA reports 83.99 tokens per second for the IQ2_XXS model with LiLiCorr in llama.cpp on a Windows 11 ARM x64 RTX Spark laptop configuration. This is a reference measurement for that configuration, not a promise for every laptop or runtime. NVIDIA’s broader design target is at least 80 tokens per second, but your result will vary with GPU, memory headroom, context, thermals, driver, runtime version and concurrency.
Deploying Bolt on embedded hardware
Use E4B for the embedded use case
NVIDIA positions Bolt E4B 4GB for fast conversational and physical agents, including local voice assistants, interactive characters and embedded interfaces. It is the variant announced for Jetson. The 74B-A9B announcement does not claim Jetson support, so do not treat a successful experiment with another runtime or an unrelated model as official support for 74B-A9B.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallE4B’s announced availability date is October 23, 2026. On October 16, 2026, that date is still in the future. A production team planning a Jetson launch should wait for the release package and confirm the final model format, supported JetPack version, runtime instructions and licensing terms rather than building a deployment around an unreleased artifact.
Rank #3
- CUTTING-EDGE PERFORMANCE – Experience next-level performance with Windows 11 Home, an Intel Core Ultra 9 Processor 275HX, and an NVIDIA GeForce RTX 5070 Ti Laptop GPU powered by the NVIDIA Blackwell architecture and featuring DLSS 4 and Max-Q technologies.
- HIGH-PERFORMANCE MEMORY AND STORAGE – Multitask seamlessly with 32GB of DDR5-5600MHz memory and store your game library on 1TB of PCIe Gen 4 SSD.
- PREMIUM ROG NEBULA DISPLAY – Immerse yourself in stunning visuals with the ultra-fast 240Hz/3ms display ideal for gaming, creation, and entertainment. Featuring a new ACR film that enhances contrast and reduces glare.
- STATE-OF-THE-ART ROG INTELLIGENT COOLING – ROG’s advanced thermals keep your system cool, quiet and comfortable. State of the art cooling equals best in class performance. Featuring an end-to-end vapor chamber, tri-fan technology and Conductonaut extreme liquid metal applied to the chipset delivers fast gameplay.
- CUSTOMIZABLE FULL-SURROUND RGB LIGHTBAR – Showcase your style with a full-surround RGB light bar that syncs with your keyboard and ROG peripherals. In professional settings, Stealth Mode turns off all lighting for a sleek, refined look.
Jetson prerequisites to plan for
NVIDIA’s Jetson guidance for its Nemotron 3.5 Lightning example documents AGX Thor and AGX Orin prerequisites. Those requirements are a useful starting checklist for an E4B design, but they are not proof that every E4B configuration will use identical versions.
- JetPack 7.2 or the version explicitly named by the final E4B guide.
- NVIDIA Container Runtime.
- Docker.
- Adequate storage for the container layers, model files, logs and any cached assets.
- Acceptance of the model-license terms before downloading or redistributing the model.
Re-check the final E4B documentation immediately before production deployment. Embedded systems have tighter thermal, storage and power limits than a workstation, so measure sustained latency and thermals in the enclosure and power mode you will ship.
Embedded validation checklist
- Load the released E4B package on the exact Jetson module and JetPack image intended for deployment.
- Verify that the model stays within the advertised 4 GB class while the application, audio pipeline, camera pipeline or device drivers are active.
- Test cold start, warm start, long conversations, dropped network connections and recovery after a container restart.
- Measure first-token latency, tokens per second, end-to-end voice latency where applicable, power draw and skin or enclosure temperature.
- Test storage exhaustion and log rotation; an otherwise healthy edge agent can fail when model caches or logs fill the device.
Can Nemotron 3.5 Bolt run directly on an iPhone or Android phone?
There is no documented direct phone deployment in the published Bolt material. NVIDIA names RTX PCs and laptops, DGX Spark, DGX Station, RTX PRO workstations and Jetson, but does not document Android, iOS, Core ML, NNAPI, Qualcomm mobile NPUs, Apple mobile NPUs or a phone-specific Bolt package.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteThat means a phone application can be a client for a Bolt service running on a supported PC, workstation or embedded device, but the published Bolt material does not establish that the model itself runs on the phone. A phone-to-local-server design should be described as remote inference from the handset’s perspective, even if the server is in the same home or office.
Rank #4
- 15.6" Full HD (1920 x 1080) widescreen LED-backlit IPS display with 165Hz Refresh Rate
- Intel Core i5-13420H Processor - up to 4.6GHz, 8 cores, 12 threads, 12MB Intel Smart Cache
- NVIDIA GeForce RTX 5050 Laptop GPU with 8GB of dedicated GDDR7 VRAM
- Massive 16GB DDR4 memory and fast 512GB PCIe Gen 4 SSD storage for accelerated load times and seamless performance.
- 1 - USB Type-C Port USB 3.2 Gen 2 (up to 10 Gbps) DisplayPort over USB Type-C, Thunderbolt 4 & USB Charging (Up to 65W)
Do not infer phone support from the 4 GB E4B label. “4 GB class” describes the model’s intended memory category, not a promise that an Android or iOS system can load it, access the required accelerator, or achieve an acceptable thermal and battery profile. Direct mobile support should only be claimed after NVIDIA publishes a phone-compatible model package and runtime path.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Choosing between laptop, workstation, Jetson and phone architectures
| Architecture | Recommended Bolt choice | What is documented | Main constraint |
|---|---|---|---|
| RTX Spark laptop or qualifying GeForce RTX PC | Bolt 74B-A9B 24GB | Local quantized inference through the listed runtimes | At least 24 GB usable GPU memory, plus sustained cooling and power |
| DGX Spark, DGX Station or RTX PRO workstation | Bolt 74B-A9B 24GB | Local or service deployment; vLLM is the named serving-oriented option | Runtime and checkpoint compatibility, concurrency and memory headroom |
| Jetson embedded device | Bolt E4B 4GB | E4B is announced for Jetson; final release documentation is pending until October 23, 2026 | Release availability, JetPack/runtime alignment, storage, power and thermal limits |
| iPhone or Android phone | No verified direct Bolt target | No official phone package or mobile-accelerator path is documented | Do not promise on-device execution; use a supported local server if a phone interface is required |
Production, customization and licensing considerations
Customization
NVIDIA describes Nemotron as an open model family intended to run from edge to cloud. NVIDIA NeMo is the customization toolkit. If you need to adapt the model rather than only run the published checkpoint, plan a separate customization workflow and verify the memory requirements of the BF16 checkpoint or any intermediate artifacts; the 24 GB local target specifically refers to the quantized deployment build.
Serving as a managed production component
NVIDIA NIM packages supported models as production-ready microservices with validated configurations and enterprise support. NIM microservices require an NVIDIA AI Enterprise license. That licensing requirement is distinct from running the open model checkpoint directly in a local runtime.
Model-license obligations
The Nemotron product information says the NVIDIA Open Model License permits use, modification, distribution and commercial deployment of the models and derivatives. Read the exact license attached to the checkpoint you select, preserve required notices and obtain any separate license required by your serving layer. A permissive model license does not automatically grant permission to use every third-party runtime, dataset, voice, image or application component in your product.
Troubleshooting common deployment failures
The model will not load on a laptop
- Check usable GPU memory after the desktop and runtime have started; a nominal 24 GB card may expose less.
- Confirm that you selected the IQ2_XXS-oriented quantized checkpoint, not BF16.
- Close other GPU applications and reduce context or batch settings temporarily to identify which allocation causes the failure.
- Verify the runtime version and model-format support. A file that loads in one backend may not be accepted by another.
Generation is much slower than the published number
- Compare your hardware and operating system with the cited Windows 11 ARM x64 RTX Spark configuration.
- Check whether the runtime is using the GPU or silently falling back to CPU layers.
- Measure with the same context length, sampling settings and concurrency as the reference test.
- Check for thermal throttling and power-limit reductions during a sustained run.
- Test LiLiCorr speculative decoding separately; it is intended to reduce latency at low concurrency, not to guarantee a gain in every workload.
A Jetson deployment fails before inference starts
- Wait for the final E4B release if you are using the pre-release period before October 23, 2026.
- Match the JetPack version, NVIDIA Container Runtime and Docker requirements in the final guide.
- Confirm adequate storage for the image and model, and verify that the model-license terms were accepted.
- Check that the device module, power mode and thermal environment match the configuration used in your validation.
A phone app cannot run the model locally
This is expected given the current documentation. There is no verified iOS or Android Bolt package. Move inference to a documented RTX, DGX or Jetson host, or wait for NVIDIA to publish a mobile deployment path; do not relabel a network client as on-device inference.
Quick Recap
A practical deployment decision
- Need a full local model on a laptop? Choose Bolt 74B-A9B 24GB only when the machine provides at least 24 GB of usable GPU memory and you can sustain its power and cooling requirements.
- Need an embedded conversational or physical agent? Plan around Bolt E4B 4GB and Jetson, but make the final decision after the October 23, 2026 release documentation is available.
- Need a phone interface? Use the phone as a client to a supported local host. Direct iPhone and Android execution is not established.
- Need customization or an enterprise service? Evaluate NeMo for customization and NIM for supported production microservices, while keeping their separate licensing and hardware requirements in your architecture record.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
