I have been playing around with lots of interesting new models and techniques these past few weeks, and somehow lost sight of an important lesson I learned last time I saw a bright shiny new capability and started venturing down rabbit holes after it. Token per second benchmarks are meaningless. Fas…

Yeah, it’s definitely a bit fiddly. I wanted to keep it completely isolated (so it isn’t messing with anything on the host) and originally used a Docker container (which was a bit simpler), but too many evals need to spawn their own docker containers, so I had to change it to a VM. It probably coul…

Thanks for the write-up! I don’t understand either the obsession with getting the fastest t/s at all cost :) and we have similar objectives it seems, apart from the fact that I’m trying to use 1 multimodal model for everything so that I can switch task without having to make any modification to the …

[image] AoE: You can drop this mod: mods/fix-qwen3.5-autoround. It’s now included in the docker. Thanks I removed that from my original post. [image] AoE: enhance your opencode setup by directly putting the model parameters in the (sub-)agent (plan, build etc.) I use it with a lot of …

Not slow, not fast ╔══════════════════════════════════════════════════════╗ ║ Benchmark: Qwen3.5 122B /w bfloat 16 — 2026-04-17 ╚══════════════════════════════════════════════════════╝ ── Sequential (1 request) ────────────────────────────── Run 1/2: [Q&A ] 256 tokens in 9.43s = …

Qwen 3.6 Here is a same-ish recipe for Qwen 3.6 35B A3B FP8. While speculative decoding is mathematically lossless, I don’t recommend this model for large tasks with lots of instructions as it is prone to token and instruction entanglement. # Recipe: Qwen3.6-35B-A3B-FP8 # Qwen3.6-122B model in FP8 …

Same 122b bfloat16 Quality w/ MTP – 38% faster Here is an updated recipe for those interested. Step 1: Update spark-vllm-docker #!/bin/bash cd ~/spark-vllm-docker git pull ./build-and-copy.sh --tf5 Step 2: Download Extended Calibration model ~ 4GiB smaller cd ~/spark-vllm-docker ./hf-download.sh…

Thanks! I’m downloading this and I’ll be testing tomorrow. I was doing A-B testing with the new Qwen3.6-35B and the 122b with MTP from Albond’s. I’ll throw this in the mix to see which one works best for me :)

@whpthomas in your Experience would you say the Qwen 3.6 122B is worse than Qwen 3.6 35B, Artificial Analysis has the 3.6 beating the 3.5 122B in almost every single benchmark. I plan to use it in opencode. What was your experience so far?

My tests (European languages) show: Qwen3.6-35B-A3B-FP8 is much much better than Qwen3.5-122B-A10B-int4-AutoRound. Don’t know of quantization (int4-autoround) makes the effect. But it’s very evident for my use cases.

Bfloat16 Quality = Speed?

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10

AoE April 17, 2026, 4:56pm 10

Agreed on fp8_e4m3. I also noticed the problem at large contexts.
For OpenCode, my suggestion is just to make this specific harness more flexible, because the Build agent should not behave like the Plan one and model tuning can help there.

NVFP4 on DGX Spark / GB10 is broken. I bought 9 of these for this feature. Requesting NVIDIA's official roadmap and response

Topic		Replies	Views
Qwen3.6-27B is out! DGX Spark / GB10 agentic-ai	300	31999	June 30, 2026
Fastest Qwen 3.5 122B Int4 recipe on DGX Spark tested and published on Spark-Arena DGX Spark / GB10 llama	59	4607	June 3, 2026
What's the best speed we can get with Qwen 3.6 27B without quantizing? DGX Spark / GB10	64	24259	July 6, 2026
Qwen/Qwen3.6-35B-A3B (and FP8) has landed DGX Spark / GB10 agentic-ai	309	30928	June 22, 2026
Qwen3.5 27B optimisation thread starting at 30+ t/s TP=1 DGX Spark / GB10 llama , agentic-ai	23	3179	May 11, 2026
Qwen3.6-27B-Dflash link DGX Spark / GB10 Projects	20	5432	April 29, 2026
DFlash LLM for DGX Spark - too good to be true? DGX Spark / GB10	37	3841	April 17, 2026
Single-Spark setups — which models do you actually run for coding, and how? (sharing mine + a test prompt) DGX Spark / GB10 spark , llama , agentic-ai , deepseek , dgx	59	6886	August 6, 2026
Qwen3.5-122B-A10B on single Spark: up to 51 tok/s (v2.1 — patches + quick-start + benchmark) DGX Spark / GB10 cuda , performance , docker , performance-tuning , llm	434	25573	June 24, 2026
DFlash for Qwen3.5-122B-A10B = 80+ tok/s on 1x Spark! DGX Spark / GB10 Projects agentic-ai	72	9042	July 27, 2026

Bfloat16 Quality = Speed?

Related topics