GLM-5.3 Flash quant showdown on 2× DGX Spark per model. Bench v6.7.1 — 76 scenarios × 2 repeats. EXL3 TR3 4bpw • TrueScore: 90.9 • Capability: 92.2 • Operational: 87.6 • Median turn: 4.30s • Long-response effective rate: ~35.4 tok/s • 53,347 total output tokens NVFP4 • Raw TrueScore: 78.5 • Capability: 71.6 • Operational: 71.9 • Median turn: 13.42s • Long-response effective rate: ~26.5 tok/s • 996,608 total output tokens EXL3 won instruction following (94.2 vs 48.9), structured output (97.2 vs 12.8), code (98.1 vs 68.5), visuals (92.3 vs 36.3), long context (100 vs 85.3) and robustness (100 vs 86.8). NVFP4 won agentic work (98.5 vs 88.0), planning (100 vs 91.2) and narrowly won safety (87.8 vs 85.6). Tool use tied at 78.4. NVFP4 completed every request with zero transport errors, but emitted 18.7× more output and took 3.12× longer per median turn. Important: this is a deployment comparison, not a pure quant-only test. EXL3 used FP8 KV, CUDA graphs and a configured 1M context ceiling. NVFP4 used Marlin, FP8 E4M3 KV, eager mode and a 262K ceiling. Verdict: EXL3 TR3 4bpw was the much better all-around deployment. NVFP4 was genuinely strong for planning and agentic workflows, but its verbosity and format behavior need fixing before it can compete on practical quality and speed.

2 DGX Sparks. Qwen3.8 Flash Next NVFP4. 64 users, independent prompts + KV caches. 32,768 output tokens in 78.82s: 415.7 tok/s Video: 1K stress. Usable-context test passed 64K/user at C16, zero leakage. Testing Spark limits, not a production recipe. @NVIDIAAI

Take two: Mac Studio m4 Max recipe Qwen3.8-Flash-Next update Average 83tok/s • MTP off→MTP3 weighted decode: 36.64→68.31 tok/s (1.86×) • MTP3→MTP6 5,999-token decode: 70.44→83.06 tok/s (+17.9%) • TrueScore: 91.8→91.9 •256k context Deployment:

GLM-5.3 Flash and Qwen3.8 Flash Next got the same uncapped prompt “build an interactive museum cube holding a living ocean that evolves continuously from calm water into a rotating supercell and hurricane, then clears back to calm. No cuts.” GLM: 30,396 reasoning / 52,182 generated tokens Qwen: 52,501 / 92,780 Clear winner is glm for this one, as Qwen’s scene needed display-only repairs to become visible and stable. No geometry, assets, or scene concepts were added. Full disclosure is on the page. Watch + interact:

Sparkbench local battle: Qwen3.8-Flash-Next (NVFP4 TP2, spec off) scored 87.0 GLM-5.3-Flash (NVFP4 TP2, DFlash2) at 85.7 This is a deployment comparison, not a universal model ranking, but Qwen was faster, more reliable, and cleaner here. The most interesting result wasn’t the score: GLM hit four ~261K-token repetition loops, while Qwen’s NEXTN path had its own collapse and had to be disabled Local inference moves fast, but serving-stack correctness matters as much as the model. Full deployment and results:

I gave GLM-5.3 Flash and Qwen3.8 Flash Next on challenge: Make one metallic sheet visibly fold into an articulated dragon, without a cut or hidden transition. It must awaken, fly, breathe particle fire, return, and unfold into the sheet. The prompt required procedural geometry, 20,000+ particles, a moonlit Japanese temple, articulated wings and tail, moving shadows, fog, bloom, orbit/zoom controls, camera presets, a live HUD, 60fps targeting, and one self-contained HTML file. GLM: 66,961 reasoning / 81,179 generated tokens Qwen: 117,329 / 153,000 Try it here

Local flash battle: GLM-5.3 Flash vs Qwen3.8 Flash Next. GLM rendered clean. Qwen needed two tiny fixes, JS declarations + a GLSL float cast GLM-5.3 Flash: ≈30.4K thinking / ≈38.7K generated Qwen3.8 Flash Next: ≈26.3K / ≈44.8K Both live:

GLM-5.3-Flash, 320B params, one DGX Spark, 256K context, 33.8 tok/s peak. Released 2 days ago Serving recipe with MTP speculative decode already up, incl. the branch pick that makes it 2x faster:

Talk about not been able to keep up, haven’t finished testing the flash version they dropped and now full 5.3 dropping in 25 hours. I’ll say it again, we’re in the inflection point of exponential growth

One DGX Spark. Qwen3.6-35B 64 users. 700+ tok/s. 32,768 tokens in 54 seconds. 38W Each user has their own prompt and their own KV cache, and vLLM batches every active stream through the GPU each step. Recipe → @NVIDIAAI