Meet FreeToken: An Edge-Native MoE Serving Engine that Runs 753B GLM-5.2 on a Single Workstation GPU
Frontier open-weight fashions are transport quicker than the {hardware} assumptions round them. Kimi-K3, GLM-5.2 and DeepSeek-V4-Flash are closing the aptitude hole with proprietary methods, however releasing parameters solely determines who can get hold of a mannequin — not who can afford to run it. Serving them nonetheless assumes datacenter-class GPU clusters, and as agentic workloads push inference demand up, that value lands hardest on particular person builders and small groups. Meanwhile, greater than a hundred million shopper machines already carry discrete GPUs. A group of researchers from UC Berkeley and UT Austin suggest FreeToken. The analysis group argued the lacking piece is just not {hardware} however a serving system: it treats a private machine as a unified, elastic inference platform slightly than a small GPU, and constantly maps computation and mannequin state onto no matter GPU, CPU, reminiscence and interconnect bandwidth the machine really has. The result’s a 35B mannequin at interactive pace on an 8 GB laptop computer GPU, 284B on a gaming desktop, and the 753B GLM-5.2 on a single workstation card.
Is it deployable?
Yes, FreeToken is Apache-2.0 on GitHub, revealed on PyPI as freetoken v0.1.2 (uv pip set up "freetoken[accel]"), and shipped as a one-click desktop app for Windows and Linux at flashml.ai. The CLI targets Linux x86_64 with an NVIDIA GPU on driver r580+ (CUDA 13). ft serve exposes OpenAI- and Anthropic-compatible endpoints on port 1919, and ft launch claude wires up Claude Code, Codex, OpenCode or OpenClaw towards your individual field.
Who it suits: solo builders, startups and SMB engineering groups whose agent token payments already exceed the price of a GPU they personal; enterprises ought to deal with it as an air-gapped or regulated-workload path, not a datacenter alternative. Strongest trade match: healthcare and authorized (knowledge by no means leaves the machine), protection, finance, and IP-heavy R&D. Typical functions: native coding brokers, personal code overview, offline contract evaluation, synthetic-data era, batch evals.
The hole it targets
Mixture-of-Experts makes native frontier inference arithmetically possible. DeepSeek-V4-Flash prompts 6 of 256 routed consultants in every of 43 layers, so solely 13B of its 284B parameters take part in any single token. Sparsity doesn’t shrink the knowledgeable pool, although — at FP4 the complete set is roughly 140 GB, so inactive consultants sit in host reminiscence and enter the execution path on demand.
The analysis group isolates three failure modes in present engines (llama.cpp, KTransformers, Ollama, MoE-Infinity):
- Prefill destroys sparsity: Thousands of tokens per layer route to almost the entire knowledgeable set, so a prefill go streams the whole pool throughout PCIe — about two seconds on an RTX 5090, 5 on PCIe 4.0 desktops, ten or extra on the x8 hyperlinks frequent in laptops.
- Static placement misses decode site visitors: llama.cpp assigns MoE tensors at load time; KTransformers pins a “sizzling” subset. Routing shifts each token, so most knowledgeable evaluations fall to the CPU whereas the GPU and the PCIe hyperlink sit idle.
- Consumer CPUs can’t carry the rest: Dual-channel DDR5 delivers 80–90 GB/s towards the 1–1.8 TB/s an RTX 4090 or 5090 attracts from on-package reminiscence.
Three mechanisms
- Bandwidth-adaptive execution (the q* coverage): Because DMA transfers and CPU knowledgeable execution learn from the identical host-memory subsystem, a saturated PCIe hyperlink leaves a residual bandwidth of
B_H − B_P. FreeToken splits every step’s m cache misses accordingly:q* ≈ m × B_P / B_Hconsultants are crammed into the GPU cache, the remainder are computed in place on the CPU, and the 2 partial sums merge precisely — no approximation, no router modification. Both bandwidths are profiled on the deployed machine (ft bench bw), which issues: measured B_P:B_H is 52.7:77.3 on an RTX 5090 server however 11.8:47.5 on a 4060 laptop computer. - Semantic-aware caching: During prefill, full-layer double buffering streams layer l+1 whereas the GPU computes layer l. Recurrent-state checkpoints are anchored at special-token boundaries — pondering blocks, software calls, software outputs — exactly the place agent harnesses truncate context, so an edit re-prefills solely the brand new suffix. During decode, a shared LRU knowledgeable cache spanning all MoE layers follows the router as a substitute of a placement frozen at load time.
- Elastic reminiscence administration: At scheduler protected factors the GPU knowledgeable cache is rebuilt underneath a revised VRAM price range with out restarting the engine or reloading the host pool. Experts are learn from disk straight into their last host structure, then pinned; no GPU warmup is required as a result of the primary request is served with a chilly cache.
Results
On an RTX 5090, FreeToken sustains 77–83 tok/s on Qwen3.6-35B-A3B (BF16) and 22–25 tok/s on DeepSeek-V4-Flash (MXFP4) — 1.5–2.3× the strongest baseline, with decode staying inside 12% of the single-turn charge throughout three agentic workloads. Worst-case TTFT stays under 44 s in each cell; llama.cpp hits 232 s, Ollama 179 s and KTransformers 946 s someplace within the matrix, previous the purpose the place agent shoppers outing.
At equal cache capability (37% of the Qwen3.6 pool), the worldwide LRU misses 16% of decode-time knowledgeable reads towards 41% for KTransformers and 62% for llama.cpp. On an 8 GB RTX 4060 laptop computer the NVFP4 construct serves 35B at 39.3 tok/s — above the 33 tok/s median decode pace measured for Codex in manufacturing traces. On a single RTX PRO 6000, GLM-5.2 (753B, 40B energetic) runs at 14.9 tok/s versus llama.cpp’s 7.3.
Data Check
| Claim | Their quantity | Independent verify | Verdict & supply |
|---|---|---|---|
| Decode, Qwen3.6-35B-A3B BF16, RTX 5090 | 77–83 tok/s | None discovered | SELF-REPORTEDPaper Fig 3 |
| Decode, DeepSeek-V4-Flash MXFP4, RTX 5090 | 22–25 tok/s | None discovered | SELF-REPORTEDPaper Fig 3 |
| Decode speedup vs strongest baseline | 1.5–2.3× | Recomputes precisely from Fig 3 | SELF-REPORTEDPaper §5.2 |
| Decode stability throughout agent workloads | inside 12% of W1 | None discovered | SELF-REPORTEDPaper §5.2 |
| Worst-case TTFT vs baselines | <44 s vs 232 / 179 / 946 s | None discovered | SELF-REPORTEDPaper §5.2 |
| 4060 laptop computer “exceeds Codex median 33 tok/s” | 39.3 vs 33 | TraceLab 33.9 is normalized; Codex pure decode median 57.1, w.avg 61.0 | MISLEADINGarXiv:2606.30560 |
| Laptop is “92% of the RTX 4090 charge” | 39.3 / 42.9 | Arithmetic appropriate, however 39.3 is NVFP4 and 42.9 is BF16 | MISLEADINGPaper Fig 5 |
| GLM-5.2 753B on one RTX PRO 6000 | 14.9 vs llama.cpp 7.3 | None discovered | SELF-REPORTEDPaper §5.3 |
| Cross-hardware lead, 5 shopper methods | 1.3–2.1× | Recomputes precisely from Fig 5 | SELF-REPORTEDPaper Fig 5 |
| Decode knowledgeable miss charge at equal capability | 16% / 39% | None discovered; hint replay, not reside serving | SELF-REPORTEDPaper Fig 4b |
| Prefill 8,192-tok chunk; overlap penalty | 1.19–1.22 s; 19/25/26% | None discovered | SELF-REPORTEDPaper Fig 4a |
| “753B on a single workstation GPU” framing | 1 GPU | True for VRAM; hosts carry 512 GiB and 192 GB DRAM | MISLEADINGPaper Table 1 |
| Baselines run at 6 CPU threads on rented servers | 6 threads | KTransformers’ core contribution is many-core AMX CPU kernels | MISLEADINGPaper §5.1 |
| “Supports greater than 20 MoE fashions” | 20+ | Public docs/fashions.md itemizes ~17 known-good MoE checkpoints | SELF-REPORTEDrepo docs |
| License and distribution | Apache-2.0, PyPI v0.1.2 | LICENSE file and PyPI JSON API each affirm | VERIFIEDGitHub, PyPI |
| Consumer discrete-GPU set up base (Steam foundation) | ~72% NVIDIA; 4060 Laptop 3.81% | Matches Valve June 2026 survey per a number of shops | VERIFIEDValve, Jul 2026 |
| Internal arithmetic throughout summary and §5 | all ratios | Every revealed ratio recomputes from Figures 3 and 5; zero errors | VERIFIEDrecomputed |
- The paper is arithmetically clear — each revealed ratio recomputes from its personal figures.
- Nothing is independently reproduced but; 9 of 16 claims are self-reported by necessity, not evasion.
- Sharpest flag: 39.3 tok/s beats Codex’s normalized 33.9, not its pure decode median of 57.1.
- “Single GPU” headlines quietly require 192–512 GB of host DRAM.
- Baseline KTransformers runs at 6 CPU threads, under the many-core AMX config it targets.
Key Takeaways
- FreeToken splits MoE cache misses between PCIe fills and CPU execution utilizing measured bandwidths, not a mounted offload rule.
- Expert output stays bit-exact — no router modifications, no knowledgeable substitution, no precision rest.
- 1.5–2.3× decode throughput over llama.cpp, Ollama and KTransformers, with tail TTFT underneath 44 s.
- 35B at 39.3 tok/s on an 8 GB laptop computer GPU; 753B GLM-5.2 on one workstation GPU.
- Apache-2.0, on PyPI and as a Windows/Linux desktop app — deployable this afternoon.
Check out the PAPER, GITHUB REPO and PROJECT. Also, be happy to observe us on Twitter and don’t overlook to affix our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Need to associate with us for selling your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar and so on.? Connect with us
The submit Meet FreeToken: An Edge-Native MoE Serving Engine that Runs 753B GLM-5.2 on a Single Workstation GPU appeared first on MarkTechPost.
