# OpenJet's Qwen3.8 Megakernel: What the RTX 3090 Benchmarks Show

OpenJet packages a faster local runtime for Qwen3.8-27B, with RTX 3090 benchmarks showing 78 tokens per second in reasoning mode and narrow compatibility limits.

By Paras Madan | 2026-10-12


OpenJet has packaged a custom CUDA engine for running Qwen3.8-27B locally. Its published RTX 3090 tests report 78 tokens per second in reasoning mode, compared with roughly 51 for llama.cpp. Those results make it worth testing for local coding if your hardware and model file match the engine's requirements.

The megakernel ships as source from version 0.4.32, released October 10. [PyPI lists version 0.4.33](https://pypi.org/project/open-jet/) on October 11. This is a runtime update for an existing model, with downloadable packages behind the performance claims.

The headline number circulating around the project is about 140 tokens per second. That comes from a particular code-generation prompt. The more useful question is what changes when the engine runs the kind of reasoning and longer-context work a coding agent actually needs.

## What OpenJet changes in local inference

OpenJet combines a local coding agent with setup and serving tools. Its [repository](https://github.com/L-Forster/open-jet) describes a workflow that checks the machine, downloads a suitable model and prepares the runtime. The agent can then edit files, run approved shell commands and resume saved sessions.

For a supported Qwen3.8-27B configuration, setup substitutes the custom engine for llama-server. Other models can continue through llama.cpp. If compilation fails, setup can fall back to llama-server and report the reason.

That fallback makes it worth checking which engine actually started before comparing speeds. A successful installation tells you the application is available; it does not by itself establish that the accelerated path is serving your requests.

## Why the megakernel is faster

The [engine documentation](https://github.com/L-Forster/open-jet/blob/master/megakernel/README.md) describes a persistent CUDA kernel that handles a complete speculative-decoding cycle. It runs the model, verifies proposed tokens and prepares further drafts on the GPU, with the host reading back accepted tokens.

Speculative decoding proposes several tokens ahead and checks whether the main model accepts them. Grouping that work can reduce overhead, but the benefit depends on how many drafts survive verification. Predictable code and less predictable prose can produce different results.

The project's llama.cpp comparison already enables MTP speculative decoding. Without speculation, it reports 42.9 tokens per second for the megakernel and 40.4 for llama.cpp. The larger gains emerge when the custom engine checks multiple drafted tokens together.

## Reading the 140-token-per-second result

The [project website](https://www.openjet.dev/) separates three measurements. The Python-function prompt reaches 139.7 tokens per second under greedy generation. Across four greedy prompts, the average is 106.7. In the reasoning-mode HTTP test, the reported rate is 78.0.

Those are different workloads, so each answers a different question. **The 140-token figure is one prompt result, rather than the average speed of the coding agent.** All these results are project-reported measurements on one RTX 3090, with one conversation at a time. They have not been independently reproduced here.

For a developer, faster generation is only part of the experience. A long input still has to be processed, tools take time to run, and a wrong edit can create another round of work. A useful local comparison would time a representative task through completion and count the corrections, alongside measuring tokens per second.

## Which hardware and model files qualify?

The custom engine requires Linux and a supported RTX 30, 40 or 50-series GPU with at least 24 GB of VRAM. CUDA 12 or newer is required, rising to 12.8 for RTX 50-series cards. The stated support is broader than the benchmark evidence: tuning and measurements were performed on a 3090.

Compatibility is also specific to the model file. The setup catalog's original Unsloth Qwen3.8-27B Q4_K_M is the supported choice. Several other Q4_K_M files are marked semi-compatible, while other quantizations use llama.cpp. Owning a GPU with enough memory therefore does not guarantee that an existing model download will use the megakernel.

## What still needs testing

The author reports close next-token agreement with llama.cpp, but the outputs are not bit-identical. Different arithmetic and rounding can change a near-tied token, after which the continuation can diverge. The [author update](https://www.reddit.com/r/LocalLLaMA/comments/1x37q7b/update_qwen_38_27b_140_toks_on_single_rtx_3090/) also acknowledges that a task-level pass-rate benchmark is still missing. Its 30K- and 96K-context agreement checks each use one sequence of repository code.

The engine is text-only and handles one sequence at a time. It ignores presence and repetition penalties, and edits to earlier context, including compaction, require processing the conversation again. Those limits can affect an agent's behaviour and latency even when decoding is faster.

**Close output distributions do not establish equivalent coding results.** The next useful evidence would be repeatable tasks on more GPUs, using the same model files and recording whether the generated changes pass their tests.

Privacy needs a similarly specific reading. Local operation works offline after setup, while the optional Slipstream mode sends planning and review work to a cloud provider. Anyone choosing OpenJet to keep code on their machine should check the selected mode and endpoints before starting a session.

I'm Paras Madan. I work at the intersection of AI engineering and distribution, helping people understand and adopt technical products.

If you're launching a developer tool, I help with technical writing, launch storytelling and founder-led distribution that explain where it works, what improves and what users should test.

[Book a 30-minute call](https://cal.com/paras-madan/30min)