🚀 v1.6.0 is out — GLM-5.3-Flash EXL3 on 2x DGX Spark 35 commits since previous release. What moved: • Decode: prose ~30.8 → ~34.8, structured 69 → ~75 @ full 7/7, essay → ~27.4 • Agents stop re-reading history: 110k replay ~98% (4s not 132s), 4x60k sessions 98.7% • Ops hardened: dual-HCA, 50ms SHM recheck baked in, prefix admission + resident tail Why not TensorFold, I know about decode speed. But I optimized it for: agentic coding, tool usage, and re-reading speed of long histories in a 1M window on a stable vLLM stack. More info here: github.com/Reederey87/glm53-…

This is why NVIDIA’s QAD quants matter for local AI. GLM-5.3-Flash-FP4-Local gets extra training to recover what compression loses. With quantization-aware distillation (QAD), the original BF16 model acts as a teacher, helping the compressed model learn to give similar answers while working with lower-precision weights. That takes more work than a regular quant conversion: NVIDIA used 500 million training tokens over 3 epochs for this release. The result is a 320B model using about 166 GiB, validated on two 128 GB DGX Sparks, with room for context. Quality stays close to BF16 on NVIDIA’s two reported benchmarks: GPQA 91.3% vs 92.1%; tool use 92.4% vs 91.9%. EXL3 at 2–3 bits per weight can squeeze models into even less memory. But fitting a bigger model doesn’t automatically mean better answers. For my coding and math work, this QAD GLM build is a much better choice than the DeepSeek-V4.1-Flash 2-3 bpw EXL3 builds I tested on the same two Sparks. I’d rather preserve more of a model’s ability than chase parameter count at the lowest possible bitrate. More releases like this, please @NVIDIAAI !!!! Model: huggingface.co/nvidia/GLM-5.… How QAD works: developer.nvidia.com/blog/ho…

I have tested nvidia/GLM-5.3-Flash-FP4-Local and one of the best DSV4.1 Flash EXL3 quads of the LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw on a 2 x DGX Spark cluster on real coding and math problems. Better output was given by the QAD NVFP4 quants of the GLM 5.3 Flash distilled by NVIDIA. No looping, better instruction following, working fine in Deepseek-Harness. It is my #1 choice for Local AI on 2 DGX Spark machines now. Thanks @NVIDIAAI for supporting local AI.

Newest cut of DeepSeek v4.1 Flash has arrived with stronger banch results.

Claude Haiku 5.5 Has Been Released Nearly a year after the prior release, Anthropic has chosen to enter the budget tier. Benchmark data indicate a clear edge over GPT-6 Luna. Costs are $0.1/$0.5 per million tokens for inputs up to 100k tokens; beyond that length, rates rise fivefold to $0.5/$2.5.

OpenAI is lowering the limits for its $200 Pro subscription. At the same time, they are reopening sign-ups for this tier. They justify the move by noting that the total output capacity will still exceed what the Pro subscription offered a month ago (prior to the Astra release) and by promising not to bring back the five-hour limits. Additionally, they promise to add new features to the subscription that won't count against usage limits. All this is happening against the backdrop of the cancelled Astra 6.1 release—due to safety concerns regarding the model—and rumors of an upcoming $500 "Pro Max" plan. Meanwhile, their main competitor, Anthropic, has just launched the highly successful Opus 5.5 and Sonnet 5.5 models.

👀

Anthropic has released Opus 5.5. According to benchmarks, it outperforms Fable 5.1; it is already available via chat, Claude Code, and the API. The price per token has been lowered compared to Opus 5—it’s now $4/$20 per million tokens. Additionally, they’ve raised the five-hour usage limits and introduced a banked reset feature. I had very mixed feelings about Opus 5; it’s the kind of model you have to constantly babysit due to issues with instruction adherence. Let's see what the feedback on 5.5 looks like.

DeepSeek V4.1 Flash leads on agentic coding and core harness benchmarks, GLM-5.3-Flash edges out general/enterprise Elo scores, and Xiaomi's MiMo-V2.6-Flash dominates security/cyber tasks.

Grok 4.7 has been released. It offers a noticeable boost while maintaining the same price and speed. Musk had previously stated that Grok 4.7 would be a new pre-trained model with 2.1 trillion parameters, but plans evidently changed. Meanwhile, training for Grok 4.8—with its 2.5 trillion parameters—was supposed to wrap up last week, and the model is likely already in the RL phase.

DeepSeek-V4.1-Flash REAP-256E running on 2× DGX Spark (GB10): 25.65 tok/s decode, 841,973-token KV pool. Fully reproducible. This is FP4, not a 3-bit or 2-bit EXL3 re-quant. What "REAP-256E" means is the routed-expert pool was pruned 384 → 256 per layer using Cerebras' REAP method — delete the experts that contribute least. Nothing was re-quantized. This is a beginning, not a finish line. More improvements planned further. Big thanks for weights to @Libertai_DAI . github.com/Reederey87/deepse…

These guys are doing an awesome project!!! I have it on my machine.

While everyone's working out how to squeeze DeepSeek V4.1 Flash — a 500B+ MoE — onto two DGX Sparks, here's a new release of the kit I actually run on mine. Two weeks of work. GLM-5.3-Flash-EXL3 on 2× DGX Spark, v1.5.0: - 1M-token context with speculative decoding live, on two desk machines - Cold prefill at 240k: 1408 tok/s across the kernel wave (grouped MoE dispatch +19.6%, fused-temp floor +13.3%,) - Decode: true +7.1% on prose; structured holds 69–70 tok/s at 1.0000 acceptance - 4.8 GiB per rank handed back to the KV pool, plus +8.1% prefill - 4 agents in flight, zero preemptions, 98.7% prefix-cache retention on repeat rounds - Acceptance stays 0.978 out to ~324k, and still 31 tok/s at 519k 2 weeks of work, more then 150 testing runs. The AB tests that reverted are written up next to the ones that shipped (this info usually lack on others' deployment kits), with numbers - drafter swap, KDA warps, fused gather, grouped-kernel occupancy — in docs/06 and docs/11, so you can follow the reasoning and qualify the development path yourself instead of trusting the README or work and try to reconcile reverted items and improve it more. Release: github.com/Reederey87/glm53-… Repo: github.com/Reederey87/glm53-…

With observed decode speed and banch result, Deepseek V4.1 Flash will lead over GLM 5.3 Flash as the best model to deploy on 2xDGX Sparks. Sincerely hope that it will be the same size or less than the 340B model.

One more!

Deepseek V4.1 Flash at API tomorrow!!!!!

Just look at the replies from some accounts—every one or two hours, day and night. A real person needs to sleep, work, and live outside X. Hermes Agent can be connected to X through xurl and the X Developer Platform, then scheduled to post and reply automatically. I have nothing against AI tools. But using bots to imitate genuine support, grow followers, and make money without being transparent is dishonest. I don’t want to waste my time talking to a bot pretending to be a person. Honestly, it’s disgusting.