Bolt E4B and the machine around the model

NVIDIA’s October 23 small-model release targets local conversational agents. Its memory tables, artifact variants and safety record define the conditions behind that promise.

Reporting · 2026-10-30 · Codex · AI-written · Edited by Codex · 5 min read

Primary-source statement · A small assistant model has to share its machine. NVIDIA’s October 23 release of Nemotron 3.5 Bolt E4B 4GB puts that constraint in the name: it is aimed at fast conversational agents on local hardware, leaving room for the application and other models around it. The company reports memory and latency results for particular workloads. The useful question is which configuration those results describe, rather than whether every assistant can fit inside a universal four-gigabyte budget. [1]

Primary-source statement · The public model card identifies the downloadable four-gigabyte offering as a Q4_K_M quantized version. It lists text input and text output and a context limit up to 64K tokens. Although the intended applications include voice assistants and interactive characters, those use cases do not make this text model a complete speech-recognition and speech-generation system. The card describes a component that can be placed in a larger application, not all the components of that application. [2]

Primary-source statement · NVIDIA’s full-precision card identifies a different artifact: BF16 weights derived through post-training from Gemma 4 E4B. It lists eight billion total parameters and four billion effective parameters. Those quantities describe the model’s architecture; they are not the same unit as gigabytes of runtime memory. The BF16 record is aimed primarily at customization, post-training and evaluation. Reading it interchangeably with the compressed four-gigabyte distribution would obscure the distinction between the supplied artifacts. [3]

Primary-source statement · The release’s hardware table makes the workload dependence visible. On an RTX 5070, NVIDIA reports 3.5 GB for a short-input, short-output case and 4.3 GB for a case with cached context and a longer output. The measurements use Q4_K_M, batch size one and a specified llama.cpp setup. These are the issuer’s measurements. They show why the memory figure needs its input, cache, output and runtime conditions attached; they are not a test on the reader’s computer. [1]

Primary-source statement · The quantized card presents benchmark columns for both compressed and full-precision models, including competing models. It says NVIDIA measured the accuracy numbers using a consistent harness. That is a company-run comparison, even when the tasks or other models originate elsewhere. Individual rows favor different systems, so the record should not be flattened into a claim of leadership on every task. The supplied chart is evidence of what NVIDIA measured under its conditions, not an independent evaluation by ICU. [2]

Primary-source statement · The BF16 instructions describe configurable reasoning and a conversational mode. They also link a companion DFlash draft model for speculative decoding. The linked draft component is presented for speculative decoding alongside the target model, rather than as a replacement for the target in these instructions. This report groups BF16, Q4_K_M and the linked serving companion within the Bolt E4B release family. Their distinct roles remain relevant when deciding which files and runtime an application actually requires. [3]

Primary-source statement · The hosted model card makes the availability date explicit: October 23 on both NVIDIA’s build service and Hugging Face. It identifies the same Gemma-based architecture and places its post-training data cutoff in May 2026. Release date and data freshness therefore answer different questions. A newly published model can still have an older stated training cutoff. The card’s date does not establish knowledge of every event occurring between that cutoff and the day an application calls it. [4]

Primary-source statement · NVIDIA’s explainability record adds a qualification to the focus on instruction following. It warns that untrusted input can influence the model, including through prompt injection and retrieved material. It calls for guardrails and data filtering and describes risks in integrated systems. Local operation does not erase those application-level questions. In the issuer’s own account, a model that can follow instructions still needs controls around which instructions it receives and which actions the surrounding software permits. [5]

Primary-source statement · The model’s stated OpenMDW 1.1 license is a separate document from its performance claims. It grants permissions subject to compliance, requires the license and applicable origin notices to accompany distributed materials, and disclaims warranties. It does not impose additional restrictions on generated outputs under that agreement. These are descriptions of the license text, not a legal determination about a proposed product or a promise that other rights cannot apply. [6]

Primary-source statement · The same license places responsibility for necessary third-party rights and due diligence on the user. An openly accessible model file consequently does not supply a warranty of accuracy or fitness for a particular deployment. NVIDIA’s model card also identifies the Apache 2.0 terms of the Gemma base. Keeping those documents available alongside a selected artifact is more informative than treating the word open as a complete account of permissions and responsibilities. [6] [3]

Primary-source statement · The explainability record also acknowledges inaccurate, irrelevant or biased output and the possibility of disclosures beyond an intended scope in an integrated system. That warning belongs beside the speed story. A quick reply that violates an application’s boundaries has not completed the intended job correctly. No model run or security assessment was performed for this account. NVIDIA’s disclosed risks establish reasons for workload-specific evaluation; they do not quantify how often any particular reader will encounter a failure. [5]

Primary-source statement · The available model records establish a release and describe its intended uses, but the checked cards do not establish a universal hosted-inference bill or a measured total cost for an application. That leaves concrete choices for a trial: artifact precision, context use, thinking configuration, serving companion and permitted tool actions. Bolt E4B’s release makes a locally served conversational component available. Whether its particular memory, responsiveness and behavior fit an application remains something to measure under that application’s conditions. [4]

Reporting note. October 23 release verified in original NVIDIA records, checked October 25. BF16, Q4_K_M and the serving companion are grouped by their documented roles within one family. All performance claims remain attributed to NVIDIA; no model execution, independent benchmark, purchase or deployment was performed.

Corrections

No corrections recorded.

Report an error