LLM Requirements
News author at LLMRequirements.com. Org byline — posts are written by the team that maintains the site.
About the publisher
LLM Requirements is a Netherlands-registered business (KVK 42058037). The same team maintains the dataset, picks, and benchmarks across the site. See About for the full publisher details and contact.
News posts cite primary sources for every load-bearing claim — vendor announcements, merged PRs, leaderboard URLs, region-local retailer pages. We don't publish unverified rumours.
Posts by LLM Requirements
-
June 2026 open-weight roundup: Gemma 4 12B, MiniMax M3, Nemotron 3 Ultra, Kimi K2.7 Code
June 2026 brought four major open-weight model launches at very different scales. Gemma 4 12B runs on a 16 GB laptop. MiniMax M3, Nemotron 3 Ultra, and Kimi K2.7 Code need multi-GPU or datacenter hardware. Here is each one, who it is for, and the primary-source benchmarks.
-
Xiaomi MiMo V2.5 on 2x DGX Spark: 32 tok/s single, 184 tok/s at 8 concurrent, 97.8 tool-quality (measured)
MiMo V2.5 is Xiaomi's 310B/15B-active omnimodal MoE, MIT open weights, that reached ~13% of OpenRouter traffic in May. We ran the NVFP4 build on our 2x DGX Spark cluster: ~32 tok/s single-stream, 184 tok/s at 8 concurrent, 1M context, 97.8/100 on a 69-scenario tool eval. Slower on deep context and concurrency.
-
Strix Halo just got up to 2.4x faster on Qwen3.6, from a llama.cpp update and no new hardware
A llama.cpp update (PR #22673, merged 2026-05-16) added built-in multi-token-prediction speculative decoding. On a Strix Halo mini-PC it lifts single-stream Qwen3.6 decode about 1.4x on the 35B-A3B MoE and about 1.8x to 2.4x on the 27B dense model, a software-only gain on a box you already own.
-
Claude Sonnet 5 vs GLM 5.2: the closed benchmark king vs the open model you can actually run
Anthropic's Claude Sonnet 5 launched June 30 at near-Opus quality, $2/$10 per million tokens intro. Open-weights GLM 5.2 lands within a point on coding and reasoning for roughly half the API cost, and unlike Sonnet 5 you can download and run it yourself. Here is the benchmark-by-benchmark comparison.
-
DeepSeek V4 Flash on 2x DGX Spark: 61 tok/s single, 261 tok/s at 16 concurrent (DSpark spec-decode, decode-only)
Reviews stop at ~40 tok/s for DeepSeek V4 Flash on a dual DGX Spark. On our 2-node cluster the DeepSeek-V4-Flash-DSpark checkpoint lifts single-stream decode to 61 tok/s and 261 tok/s at 16 concurrent. The lever is speculative decoding, not FP4, and the kernel still crashes under concurrent long context.
-
DGX Spark runs Qwen3.5-122B at 59 tok/s general and 81 on agent traffic, with speculative decoding (not NVFP4)
On a single DGX Spark, Qwen3.5-122B-A10B with DFlash block-speculative decoding runs about 59 tok/s on general decode and about 81 tok/s on real agent and tool-call traffic. For agentic work the lever is speculative decoding, not the stock low-50s decode reviews quote, and not NVFP4 (no measured win on this model).
-
NVIDIA announces Nemotron 3 Ultra — 500B/50B open MoE, weights expected this week
Jensen Huang unveiled Nemotron 3 at Computex on 2026-06-01. The Ultra tier is 500B/50B-active MoE, NVFP4-trained on Blackwell. AA scores it Intelligence Index 48 — top of US open-weights, ahead of Gemma 4 31B, Nemotron 3 Super, gpt-oss-120B. Weights not yet on HF; this is announcement-only.
-
NVFP4 + Qwen 3.6 35B-A3B on DGX Spark: 97 tok/s single, 322 tok/s at 8 concurrent (decode-only)
NVIDIA's official NVFP4 checkpoint of Qwen 3.6 35B-A3B (May 28). On a DGX Spark running the technigmaai/dgx-spark recipe we measured 97 tok/s single-stream and 322 tok/s aggregate at 8 concurrent with prefix-cached prefill (42 tok/s per stream). Community MTP-3 sweet spot adds another lift on FP8.
-
RTX 5090 UK in-stock floor drops to £2,990 at Scan — first softening since March's DRAM spike
Scan UK's lowest in-stock RTX 5090 (32 GB) is now £2,989.98 for the ZOTAC SOLID OC, marked down from £3,118.99 with next-day dispatch. A cheaper Gigabyte WINDFORCE OC sits at £2,879.99 but pre-order only with no delivery date. The first material UK 5090 softening since March's DRAM contract spike.
-
Framework lifts Strix Halo 128 GB another 5.3 % in UK and EU — £3,159 / €3,559
Framework's Desktop DIY AMD AI Max+ 395 128 GB jumped £2,999 → £3,159 (UK) and €3,379 → €3,559 (EU) on the configurator this week — the second EU/UK hike in 11 days after the May 16 +12 %/+11 % move. Stock remains pre-order only across both regions.
-
DRAM shortage: RTX 5090 +29%, 4090 +28%, used 3090 +17%, GB10 mini-PCs +34% in May
The DDR5 / GDDR6 / GDDR7 contract spike that started in March is now visible at retail across the LLM-rig segment. 16 builds repriced this week — RTX 5090 cards now $3,658+ over the $1,999 MSRP, 4090 partner stock dried up, used 3090 / AMD W7900 / GB10 mini-PCs all moved between +6% and +34%.
-
Cursor Composer 2.5 ships on open-weights Kimi K2.5 at $0.50/M input
Cursor's May 18 Composer 2.5 release keeps Composer 2's Kimi K2.5 backbone and bets RL fine-tuning closes the gap to Opus 4.7 on coding benchmarks. Standard tier is $0.50/M input + $2.50/M output — roughly a tenth of frontier closed-ref pricing. The base weights are public under Modified MIT.
-
GB10 mini-PC pricing diverges: DGX Spark drops 15% in the UK, MSI EdgeXpert holds at €4,700 in France
CyberPowerPC UK cut the NVIDIA DGX Spark 128 GB from £4,699 to £3,999 (−14.9%) in mid-May; LDLC France still lists the equivalent MSI EdgeXpert MS-C931 (4 TB) at €4,699.95. UK buyers are now meaningfully cheaper than EU on the same GB10 silicon.
-
Multi-token prediction lands in llama.cpp (beta)
PR #22673 merged MTP support into llama.cpp on May 16 2026. A community benchmark on Qwen 3.6-27B reported decode speed jumping from 38 to 65 tok/s on a single RTX 5090 — a 1.71× speedup.
-
Framework lifts Strix Halo 128 GB prices 12% in EU, 11% in UK
Framework's Desktop AMD Ryzen AI Max+ 395 128 GB SKU jumped from €3,029 to €3,379 in the EU and £2,699 to £2,999 in the UK mid-May, on continuing DDR5 module shortage. US pricing held.