Blackwellboy / Work
The strongest public deployments and studies from the lab. Results come first; the exact lane, conditions and failure boundaries stay attached rather than being flattened into one benchmark table.
Large models where the interesting work is the system around the weights: topology, kernels, memory, speculative decoding, serving and the point where a nominally healthy lane actually fails.
The public bring-up package documents the Neuron IQ1_S GGUF running end to end across three and four Sparks, including local-NVMe loading, NCCL fabric receipts and the TP geometry. The 3-Spark path uses 512/512/512 FFN-width shards; the recipe also records why raw token-ID speed prompts are not quality tests.
The model loads on the tempting b12x path and even returns healthy API status, then the first real request kills the engine because the TMA attention kernel cannot JIT on the stock CUDA-13.0 driver. The working answer is a different lane, not driver surgery.
Dual-Spark Hy3 serving with NVFP4 W4A16, native MTP, FP8 KV cache and structured tool calling. The campaign measures both speed and an accuracy slice, because a quant that serves quickly but damages the model is not a successful deployment.
A live TP2 lane became a systems study: MTP acceptance by task family, reasoning-field behaviour, cache effects and long-context retrieval. The server advertises a 1M window, but the checkpoint is trained to 65,536 and cold retrieval becomes unreliable rather than failing cleanly.
More model history lives on the Models page. It separates full campaigns from targeted tests and control models so breadth does not get confused with depth.
The failures are often more reusable than the headline speed number.
Weights load, KV allocates and /health returns 200. The first inference triggers a CUDA-13.2 TMA JIT path the stock 13.0 driver cannot build. A green health endpoint was not a serving proof.
The server accepts the window and never raises an error, yet cold retrieval becomes unreliable from much shallower depths. Process survival and useful long-context behaviour are different claims.
Read the model studyA published HumanEval+ advantage used different sampling settings between the arms. Once the sampling was matched, the gain disappeared and the sign reversed.
Read the replicationIn the tested OpenAI-style path, prior assistant reasoning was not resent. The template rendered empty think blocks into history and the multi-turn gate collapsed. Preserving reasoning restored firing.
Read the mechanismOn the Laguna serving sweep, deeper speculative settings became slower. Draft depth is a verifier-economics problem, not a knob that monotonically buys throughput.
Read the sweepControlled studies where the result mattered more than whether it agreed with the starting hypothesis.
Reasoning behaviour
Thinking was not controlled by a flag or prompt length alone. A dense ten-rule persona block fired 3/40, while a much longer full agent prompt fired 24/40. Summarization fired 0/105; the same grid on Qwen 3.6 35B-A3B fired 400/400. The useful result is a prompt × task gate, not a simple dose-response story.
400 grid turns, 10 conditions by 4 task types, n=10 per cell, 450 turns logged including parser and criteria phases, 0 HTTP errors, single-turn. Laguna S 2.1 NVFP4 rev 0761412 under vLLM 0.25.1 on one DGX Spark GB10. These are within-run cells on this build and do not generalise across quants or runtimes.
Benchmark replication
A published “thinking-on wins” result had the two arms running different sampling settings. With sampling matched, the gain disappeared and the sign reversed: ON 89.84, OFF 90.85. Thinking-on was steadier across seeds, but cost roughly ten times the wall clock: 200.2 s mean per problem versus 18.5 s.
984 requests, 164 problems by 2 arms by 3 seeds, arms interleaved, identical sampling both arms at temperature 0.7 / top_p 0.95 / top_k 20, max_tokens 12,288, 0 transport errors, evalplus 0.3.1. NVFP4 rev 0761412 under vLLM 0.25.1 on one GB10.
Multi-turn serving
In the OpenAI-style client path tested, prior assistant reasoning was not resent. The template then rendered earlier assistant turns with empty think blocks and the model stopped firing its reasoning gate. Identical transcripts fired 10/10 with reasoning preserved and 0/10 stripped at both depths tested.
Depth 10 and depth 20 conversations, identical transcripts both arms, max_tokens 4096. Measured on the 0xSero 3.25bpw EXL3 hybrid lane. This is a mechanism result for the tested client/template path, not a claim that every OpenAI-compatible client strips reasoning.
Hypothesis test
This was run expecting to confirm a community hypothesis that long system prompts suppress thinking by displacing the trained identity. It failed the direct test: moving identity to the literal first line took firing from 5/40 to 0/40. Neutral and topical filler reopened the gate by similar amounts, pointing away from identity itself.
n=40 per cell, arms interleaved within each run, two builds. NVFP4 bare 2/40 against identity 17/40, neutral filler 10/40, topical filler 11/40. The 3.25bpw hybrid bare 0/40 against 13/40, 14/40, 10/40. Qwen 3.6 35B-A3B null control fired 480/480.
Performance tuning
Speculative depth beyond K=7 made this model slower at every concurrency level tested. The single-stream winner reached 26.2 tok/s median and 61.65 tok/s aggregate at concurrency 4. The grid’s best aggregate was a different cell: K=6 with max-num-seqs 8 at 62.26 tok/s.
20 cells, K 5 to 9 by max-num-seqs 4 to 32, single run per cell. NVFP4 rev 0761412 under vLLM 0.25.1 on one GB10, Hermes bench v1 short subset. Every losing cell is published.
Production stability
3,096 HTTP-200 turns across 409 sessions, with zero crashes and zero restarts, inside a real agent pipeline. Thinking fired on only 3 turns despite being requested every turn; the later reasoning-stripping study explains that anomaly.
One 12-hour run, production profile K=7 and max-num-seqs=32, max_tokens 8192. 3,099 rows logged; 3,096 returned HTTP 200. Correction 2026-07-25: 409 sessions = 400 completed + 9 killed at the driver token cap; the three extra lines are integrity probes and all returned HTTP 200.
The website tells the story; GitHub holds the runnable detail.