Open models with serving tuned on real hardware. Buy once, run anywhere.
Core plan required
We keep re-optimizing and re-benchmarking as engines move, and you always download the latest verified version. Cancel anytime. Nothing is charged until you confirm at checkout.
RunInfraby RightNow© 2026 RunInfra. All rights reserved.
Single-stream decode with one variable between the two columns of this comparison.
Both columns were measured on 2026-08-01 in one window, one container, one engine install, on the same 4x B200 cards, run back to back with an idle VRAM drain between them, and the comparison tool verified the request profiles identical across the arms before rendering any number.
The baseline column is this same configuration with speculative decoding off, not a different engine and not a failed serve.
The optimized column enables the DSpark speculative decoding that ships fused inside the DeepSeek-V4-Flash-0731 checkpoint, which the released engine does not enable on its own.
The throughput figures are per-request decode rates at concurrency 1, the speed one user experiences, not per-GPU figures and not node aggregates: 113.37 output tokens per second p50 with speculation off and 362.86 with it on.
The latency stats beside each column are end-to-end request times from the same runs: the optimized median request finishes in 3.1 seconds instead of 9.2, and its time to first token is 232.93 ms against 220.21, a cost of about 13 ms.
The profile, stated exactly: single-stream sequential, concurrency 1, 8 timed requests per arm, 1024 prompt tokens target, 1024 output tokens requested per request, temperature 0.0, ignore_eos on, streaming, a fixed seed, and 8 of 8 requests ok on each arm.
Know the boundary of this result before you deploy on it: the advantage is a low-concurrency fact, and the same window measured how it narrows with load on 32768-token-prompt workloads, plus 168 percent per GPU at concurrency 1, plus 33.6 percent at concurrency 8, and minus 6.9 percent at concurrency 32 where the median end-to-end time stays flat.
The optimized configuration also ran in the measurement slot that this campaign's A/A control series showed to be systematically disadvantaged, so the comparison understates the effect rather than overstating it.
Workload profile 32768 prompt tokens and 2048 output tokens per request, repo-scale coding shape, closed loop, 8, 16 and 64 timed requests at concurrency 1, 8 and 32, temperature 0.0, fixed seed, measured on the optimized configuration in the same one-container window as the headline pair.
Deployment guidance: request concurrency is a knob you set on your own deployment, and each point states the measured trade at that setting, not a package speedup.
Deployment guidance, not a speedup claim: the concurrency knob belongs to the buyer. The same window measured the speculation advantage narrowing with load, and at concurrency 32 the optimized configuration's aggregate throughput sat 6.9 percent BELOW its own no-speculation baseline while median end-to-end time stayed flat, with the optimized arm in the measurement slot the A/A control series showed to be disadvantaged. Batch-first deployments above concurrency 8 should benchmark their own operating point; interactive and low-concurrency serving is where this recipe earns its multiplier.
Run the model on your infrastructure, without deprecations, silent swaps, or provider restrictions.
Buy once instead of paying per token; your deployment cost follows the GPUs you run.
Inference runs on your infrastructure, keeping prompts and outputs on your network.
Delivery contract
Download
./models/kit.zipArchive root
deepseek-v4-flash-b200x4-v4turbo/Model directory
deepseek-v4-flash-b200x4-v4turbo/weights/Filled by --weights
Kit contains
Fetched separately
f9304227d189...6d67a51f3bfcThe checkpoint ships its own trained draft module fused into the weights; the released engine leaves it off. The draft module is DeepSeek's work; the enablement, configuration and measured operating envelope are this package's.
Data-parallel 4 with expert parallelism, FP8 KV cache, 256-token attention blocks, the model-specific tokenizer and reasoning parser flags the fused checkpoint requires, an engine readiness timeout sized to this model's real cold-start, and a 131072-token context sized to the measured workloads. Every flag carries its reason in the shipped configuration notes.
Measured accuracy
Output parity by construction
Accuracy recovery not measured
This package does not publish both a baseline and an optimized score on this metric, so there is no recovery figure to state.
Methodology
Two licenses apply: the RunInfra package license you purchase under, and the base model's own open-source license.
RunInfra Package License v1.1 (2026-07-25)
Buy once. No meter and no expiry. No RunInfra service is required at runtime. The full terms ship inside the kit.
This package is licensed to the purchasing workspace, one time, for the exact version purchased. Everyone in that workspace may run it in production without limits: unlimited inference, on any hardware or cloud the workspace controls, for any lawful commercial purpose. You may modify the configuration and tooling for your own use.
Everything the model produces for you belongs to you. You may use, sell, and build products on the model's outputs without restriction or royalty. Serving the model to your own customers as part of your product is use, not redistribution, and is fully allowed.
You may not redistribute, resell, sublicense, rent, publish, or otherwise make the package or its artifacts available to any third party. That covers the serving configuration, scripts, kit archive, and benchmark receipts, whole or in part, modified or not. The upstream model weights are not included in this kit. You fetch them separately from Hugging Face at the pinned revision, and they remain governed by the model's own license, not this package license. The license belongs to the purchasing workspace and cannot be transferred separately from it.
The underlying model remains under its upstream open-source license, which is included in this kit with attribution and a statement of RunInfra's modifications. Nothing in this license restricts rights the upstream license grants you for the ORIGINAL model; the restrictions above apply to RunInfra's optimized package.
You purchased this exact version, proven against the exact engine version named in the receipt. It stays downloadable to your workspace and never expires. The benchmark receipt describes measurements taken at verification time on the named hardware; RunInfra does not promise future updates to this version, and later package versions are separate purchases.
If the workspace redistributes the package or its artifacts, this license terminates for that workspace. Sections about your outputs survive termination for outputs already produced.
MIT, as published by the model author on Hugging Face. The full license text ships inside the kit.
Company
RunInfra is a sub-product of RightNow Research Lab.
Y CombinatorBacked by Y Combinator.
NVIDIA InceptionMember of NVIDIA Inception.
