MiniCPM5-2B-DSpark vs Granite 4.2 3B: Two Quiet Apache-2.0 Drops for Small, Fast, Self-Hosted Serving

August and early September 2026 produced two quiet, Apache-2.0 releases aimed at the same job — small models you serve yourself — and they went about it from opposite ends. Granite 4.2 3B is IBM's laptop-sized reasoning model, whose weights landed on Hugging Face in early August and whose model card and technical blog went out on August 25, 2026. MiniCPM5-2B-DSpark is not a model at all in the same sense: it is a 323.8M-parameter DSpark draft checkpoint that OpenBMB posted on September 6, 2026 without any announcement, built to make MiniCPM5-2B — the 2.52B dense on-device model ModelBest announced at WAIC on July 19, 2026 — generate tokens faster through speculative decoding. One release is a standalone small reasoning model; the other is an acceleration accessory for a small model. Both sit under Apache-2.0, both are self-host only, and neither has been touched by an independent benchmark yet.

Strip the marketing layer off and the practical question a team faces is straightforward: for a small, self-hosted serving workload in late 2026, do you want IBM's 3B reasoning specialist or ModelBest's 2B agent-oriented stack with a free draft bolted on? The evidence is thinner than either vendor's spec sheet suggests, so this article walks the release facts, the one set of same-suite numbers that actually exists, and the workload difference that should drive the choice.

The two releases, what is knowable

Granite 4.2 3B is IBM's smallest reasoning model, post-trained from Granite-4.1-3B-Base. It is dense and text-only — 40 layers, grouped-query attention, a 128K native context that extends to 512K in a fifth pre-training phase, and three reasoning modes: full thinking by default, a low-effort mode, and a non-thinking path. IBM's card is explicit that the 3B deliberately skipped the agentic reinforcement-learning block the larger Granite 4.2 siblings received, so it lists no SWE-Bench results and is a reasoning model rather than an agent model. It serves through vLLM, SGLang, Transformers, GGUF, and Ollama (granite4.2:3b), and it has no hosted API. Its headline card numbers — 78.33 on AIME 2025, 54.80 GPQA, 69.71 LiveCodeBench v6 — are vendor-reported and unreproduced as of today.

The ibm-granite/granite-4.2-3b card above is the public face of that release: a reasoning-tuned 3B with a long-context story and no agentic claims.

MiniCPM5-2B-DSpark, by contrast, exists only in service of its target. The draft is five full-attention layers at 323,776,001 parameters with a block size of seven candidate tokens per forward pass, trained for six epochs on 7.05 billion tokens of MiniCPM5-2B-generated responses. Its repo reports acceptance length — 5.52 accepted tokens per verification step on aggregate at greedy decoding, 6.11 on code — but no tokens-per-second figure. The target it accelerates, MiniCPM5-2B, is the more interesting product: a dense 2.52B, 42-layer, 128K-context model whose launch materials emphasize agentic capability — 200B-scale agent mid-training, a large agent-SFT set, and agent RL alignment, per ModelBest's July announcement — plus native long-context and hybrid fast/deliberate modes. On ModelBest's own comparison suite the target scores an average of 53.9 against same-class open models. Every one of those numbers is the vendor's.

The openbmb/MiniCPM5-2B-DSpark card above is the entire release surface: model card, config, one Safetensors file, and no announcement.

The only same-suite comparison that exists

Cross-vendor scoreboards are mostly apples-to-oranges, but there is one place where Granite 4.2 3B and MiniCPM5-2B were measured together: ModelBest's own evaluation table for MiniCPM5-2B, which lists granite-4.2-3B among its baselines. On that suite, MiniCPM5-2B averages 53.9 against Granite 4.2 3B's 42.7. Read that number exactly as what it is — one vendor running both models on one suite, unreproduced, and selected to make its own model look good. It is not a verdict. What it does tell you is that ModelBest was confident enough to put a competitor's 3B on the card, and the gap it claims is largest exactly where its own training invested: long-context and agentic rows. Take it as a directional claim, verify it on your own data, and weigh IBM's own separate card claims before you decide anything on it.

The scoreboard, honestly labeled

• What ships — MiniCPM5-2B-DSpark: a 324M draft for MiniCPM5-2B (2.52B dense target), not standalone. Granite 4.2 3B: a standalone ~3B dense reasoning model.

• Role — MiniCPM5-2B stack: on-device agent and tool-use emphasis, per ModelBest. Granite 4.2 3B: reasoning specialist, deliberately non-agentic.

• Context — MiniCPM5-2B target: 128K native. Granite 4.2 3B: 128K native, extending to 512K.

• Acceleration — MiniCPM5-2B-DSpark: external DSpark draft, seven tokens per pass, greedy acceptance ~5.5 per step (repo-reported). Granite 4.2 3B: none shipped with this release.

• License — both Apache-2.0.

• Evidence — MiniCPM figures are ModelBest card claims (its suite: 53.9 vs Granite 42.7); Granite figures are IBM card claims (AIME 2025 78.33, GPQA 54.80). No independent run of either exists.

The scoreboard above carries the same sourcing as the prose: every figure on it is vendor-reported, and the one same-suite number either vendor has published is ModelBest's own.

The difference that outranks every benchmark

Before the scores, decide which kind of small model your workload needs, because the two releases answer different questions. Granite 4.2 3B is built for reasoning and long context on constrained hardware: a 128K native window that stretches to 512K, three per-query thinking modes, a footprint that runs on a laptop, and an explicit decision to skip agentic training. If your job is long-document analysis, repository-scale context, or reliable reasoning on a small box, that is a coherent fit. MiniCPM5-2B is built for the opposite emphasis: agentic behavior and tool use on device — the draft existing at all signals that ModelBest expects this target to be served interactively and wants the tokens per second back. If your job is an on-device agent loop that calls tools and holds long conversations, that is the stack aimed at you.

The draft changes the economics of that second choice in a way Granite's release does not address. MiniCPM5-2B is dense, and speculative decoding pays off most in exactly the dense, memory-bound regime — a small fixed model proposing tokens to a slightly larger fixed one. An external drafter that adds roughly 650MB of BF16 weights to a ~5GB target, with a documented SGLang serving path and a greedy acceptance rate around five-and-a-half tokens per verification step, is a credible throughput lever for a model you serve at single-request concurrency. But the caveat is unavoidable: OpenBMB has published no end-to-end speedup, so the lever is uncalibrated. IBM ships no equivalent external drafter for Granite 4.2 3B as part of this release — you run it dense, and its speed story is simply that the 3B is small.

Who should run which

• Run Granite 4.2 3B if your workload is long-context reasoning on a laptop-class box — documents, repositories, deep single-threaded analysis — and you want three thinking modes and a 512K ceiling on an Apache-2.0 model you fully control. Accept that there is no agentic training behind it and no hosted API.

• Run the MiniCPM5-2B stack with its DSpark draft if your workload is an interactive, tool-calling on-device agent where the target fits your memory budget and you want the throughput a draft can buy back. Budget time to measure the draft's real speedup on your own SGLang build, because no number for it has been published.

• Run both behind a routing layer if you are genuinely unsure. Neither model is in a hosted catalog today, OrcaRouter's included — both are self-host propositions — but a router that fronts your own endpoints with automatic failover lets the new draft stack sit on a test path while production stays on the proven one, and its 0% markup pass-through on the hosted models around them keeps the A/B comparison cheap. The point is reversibility: swap a model name, measure, and switch back if the one-day-old checkpoint stalls.

What to watch next

Two things will settle this matchup faster than any opinion. First, an independent run of MiniCPM5-2B that confirms or sinks the 53.9 average and the agentic claims — a 2B scoring that well on SWE-Bench-style tasks would be a genuinely new data point for the on-device class. Second, IBM's own reasoning-model numbers surviving community reproduction; the 3B's 78.33 AIME 2025 is the kind of claim that moves when someone else runs it. Until one of those lands, the honest position is: Granite 4.2 3B is the safer, better-documented small model today, and the MiniCPM5-2B stack is the more interesting bet — a faster-on-device agent if its claims hold, with a draft whose payoff even its own vendor has not yet measured.